🧠 Prism MCP — The Mind Palace for AI Agents

Your AI agent forgets everything between sessions. Prism fixes that — then teaches it to think.

Prism v11.5.1 is a true Cognitive Architecture inspired by human brain mechanics. Beyond flat vector search, your agent now forms principles from experience, follows causal trains of thought, and possesses the self-awareness to know when it lacks information. Your agents don't just remember; they learn. With v11.5.1, the entire cognitive pipeline — including ledger compaction, task routing, semantic search, and the new Deep Research Intelligence — runs 100% on-device or via secure clinical discovery (PubMed/ERIC), backed by prism-coder:7b, a HIPAA-hardened local LLM. No API keys for core features. No data leaves your machine.

npx -y prism-mcp-server

Works with Claude Desktop · Claude Code · Cursor · Windsurf · Cline · Gemini · Antigravity — any MCP client.

https://github.com/dcostenco/prism-mcp/raw/main/docs/prism_mcp_demo.mp4

• 🔬 v11.5.1 Deep Research Intelligence (Auto-Scholar)
• ⚡ Zero-Search Retrieval (HRR Architecture)
• Why Prism?
• Quick Start
• The Magic Moment
• Setup Guides
• Universal Import: Bring Your History
• What Makes Prism Different
• Cognitive Architecture (v7.8)
• Data Privacy & Egress
• Use Cases
• What's New
• How Prism Compares
• CLI Reference
• Tool Reference
• Environment Variables
• Architecture
• Scientific Foundation
• Milestones & Roadmap
• Troubleshooting FAQ

Prism v11.5.1 transforms your AI agent from a "Coder" into a "Clinical Scientist." It features a Tavily-Enhanced Multi-Provider Discovery Pipeline that grounds Gemini 2.5 Flash's thinking in real-world empirical data.

Prism features a cutting-edge Zero-Search Retrieval system for its cognitive memory, moving beyond traditional vector databases toward a mathematically direct, $O(1)$ retrieval model.

Zero-Search Retrieval uses Holographic Reduced Representations (HRR) to "ask the vector" directly. Structured facts are bound into a single, high-dimensional "superposition" vector (typically 4096-dims). Retrieval is a direct mathematical unbinding operation (circular correlation).

1. Tavily AI (Elite): Primary discovery engine for AI-native deep crawling and PDF/Abstract extraction.
2. PubMed (NCBI) (Clinical): The world's largest biomedical database for clinical citations.
3. ERIC (Education Research) (Behavioral): The definitive database for ABA and pediatric interventions.
4. Semantic Scholar (Academic): AI-powered research tool providing "TLDR" summaries of 200M+ papers.
5. DuckDuckGo Lite (Fallback): Privacy-focused web discovery for general context.

Synalux is a high-compliance, local-first Practice Management System for ABA and Pediatrics. It is the flagship implementation of the Prism v11.5.1 engine, utilizing Zero-Search Retrieval and Parallel Academic Discovery to provide clinicians with real-time, evidence-based reasoning.

Every time you start a new conversation with an AI coding assistant, it starts from scratch. You re-explain your architecture, re-describe your decisions, re-list your TODOs. Hours of context — gone.

Prism gives your agent a brain that persists — and then teaches it to reason. Save what matters at the end of each session. Load it back instantly on the next one. But Prism goes far beyond storage: it consolidates raw experience into lasting principles, traverses causal chains to surface root causes, and knows when to say "I don't know."

📌 Terminology: Throughout this doc, "Prism" refers to the MCP server and cognitive memory engine. "Mind Palace" refers to the visual dashboard UI at localhost:3000 — your window into the agent's brain. They work together; the dashboard is optional.

Prism has three pillars:

1. 🧠 Cognitive Memory ($O(1)$ Zero-Search) — Prism uses Holographic Reduced Representations (HRR) to eliminate "searching" entirely. Memories are unbound mathematically from a superposition vector in constant time ($O(1)$), regardless of library size. Re-ranking is powered by the ACT-R model, mimicking biological recency and frequency.

2. 🔗 Multi-Hop Causal Reasoning — Prism doesn't just find "similar" things. Spreading activation traverses the causal graph and brings back context connected to your current problem through logical "trains of thought."

3. 🏭 Autonomous Execution (Dark Factory) — When you're ready, Prism can run coding tasks end-to-end with a fail-closed pipeline where an adversarial evaluator catches bugs the generator missed — before you ever see the PR. (See Dark Factory.)

• Node.js v18+ (v20 LTS recommended; v23.x has known npx quirk)
• Any MCP-compatible client (Claude Desktop, Cursor, Windsurf, Cline, etc.)
• No API keys required for core features (see Capability Matrix)

Add to your MCP client config (claude_desktop_config.json, .cursor/mcp.json, etc.):

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"]
    }
  }
}

⚠️ Windows / Restricted Shells: If your MCP client complains that npx is not found, use the absolute path to your node binary (e.g. C:\Program Files\nodejs\npx.cmd).

That's it. Restart your client. All tools are available. The Mind Palace Dashboard (the visual UI for your agent's brain) starts automatically at http://localhost:3000. You don't need to keep a tab open — the dashboard runs in the background and the MCP tools work with or without it.

🔮 Pro Tip: Once installed, open http://localhost:3000 in your browser to view the Mind Palace Dashboard — a beautiful, real-time UI of your agent's brain. Explore the Knowledge Graph, Intent Health gauges, and Session Ledger.

🔄 Updating Prism: npx -y caches the package locally. To force an update to the latest version, restart your MCP client — npx -y will fetch the newest release automatically. If you're stuck on a stale version, run npx clear-npx-cache (or npm cache clean --force) before restarting.

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"],
      "env": { "PRISM_DASHBOARD_PORT": "3001" }
    }
  }
}

🔑 The core Mind Palace works 100% offline with zero API keys — including semantic vector search with embedding_provider=local. Cloud keys unlock text generation features (Briefings, compaction, pipelines). See Environment Variables.

💰 API Cost Note: With embedding_provider=local, semantic search is fully free and offline. Cloud providers (GOOGLE_API_KEY for Gemini, VOYAGE_API_KEY, OPENAI_API_KEY) have generous free tiers. BRAVE_API_KEY offers 2,000 free searches/month. FIRECRAWL_API_KEY has a free plan with 500 credits. For typical solo development, expect $0/month on the free tiers.

Session 1 (Monday evening):

You: "Analyze this auth architecture and plan the OAuth migration."
Agent: *deep analysis, decisions, TODO list*
Agent: session_save_ledger → session_save_handoff ✅

Session 2 (Tuesday morning — new conversation, new context window):

Agent: session_load_context → "Welcome back! Yesterday we decided to use PKCE
       flow with refresh tokens. 3 TODOs remain: migrate the user table,
       update the middleware, and write integration tests."
You: "Pick up where we left off."

Your agent remembers everything. No re-uploading files. No re-explaining decisions.

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"]
    }
  }
}

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"]
    }
  }
}

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"]
    }
  }
}

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"],
      "env": {
        "PRISM_STORAGE": "local",
        "BRAVE_API_KEY": "your-brave-api-key"
      }
    }
  }
}

Always start the conversation by calling `mcp__prism-mcp__session_load_context(project='my-project', level='deep')`.
When wrapping up, always call `mcp__prism-mcp__session_save_ledger` and `mcp__prism-mcp__session_save_handoff`.

# Load context before a session
prism load my-project --level deep

# Machine-readable JSON for parsing in scripts
prism load my-project --level deep --json

# From a shell or run_command tool
prism load my-project --level standard --json

# Or via the wrapper script
bash ~/.gemini/antigravity/scratch/prism_session_loader.sh my-project

# Quick check — human-readable
prism load my-project

# Parse JSON in scripts
CONTEXT=$(prism load my-project --level quick --json)
SUMMARY=$(echo "$CONTEXT" | jq -r '.handoff[0].last_summary')
VERSION=$(echo "$CONTEXT" | jq -r '.handoff[0].version')
echo "Project at v$VERSION: $SUMMARY"

# Explicit storage backend (v9.2.2 — prevents split-brain)
prism load my-project --storage local --json
prism load my-project --storage supabase --json

# Role-scoped loading
prism load my-project --role qa --json

# Use in CI/CD to verify context exists before deploying
if ! prism load my-project --level quick --json | jq -e '.handoff[0].version' > /dev/null 2>&1; then
  echo "No Prism context found — skipping context-aware deploy"
fi

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"],
      "env": {
        "PRISM_STORAGE": "supabase",
        "SUPABASE_URL": "https://your-project.supabase.co",
        "SUPABASE_KEY": "your-supabase-anon-or-service-key"
      }
    }
  }
}

# Via CLI (requires supabase CLI + project linked)
supabase db push

# Or apply a single migration via the Supabase dashboard SQL Editor
# Paste the contents of supabase/migrations/0NN_*.sql and click Run

git clone https://github.com/dcostenco/prism-mcp.git
cd prism-mcp && npm install && npm run build

{
  "mcpServers": {
    "prism-mcp": {
      "command": "node",
      "args": ["/path/to/prism-mcp/dist/server.js"],
      "env": {
        "BRAVE_API_KEY": "your-key"
      }
    }
  }
}

{
  "mcpServers": {
    "prism-mcp-cloud": {
      "command": "npx",
      "args": ["-y", "supergateway", "--url", "https://your-prism-app.onrender.com/sse"]
    }
  }
}

❌ Don't use npm install -g: Hardcoding the binary path (e.g. /opt/homebrew/Cellar/node/23.x/bin/prism-mcp-server) is tied to a specific Node.js version — when Node updates, the path silently breaks.

✅ Always use npx instead:

{
  "mcpServers": {
    "prism-mcp": {
      "command": "npx",
      "args": ["-y", "prism-mcp-server"]
    }
  }
}

npx resolves the correct binary automatically, always fetches the latest version, and works identically on macOS, Linux, and Windows. Already installed globally? Run npm uninstall -g prism-mcp-server first.

❓ Seeing warnings about missing API keys on startup? That's expected and not an error. API key warnings are informational only — core session memory and semantic search (with embedding_provider=local) work with zero keys. See Environment Variables for what each key unlocks.

💡 Do agents auto-load Prism? Agents using Cursor, Windsurf, or other MCP clients will see the session_load_context tool automatically, but may not call it unprompted. Add this to your project's .cursorrules (or equivalent system prompt) to guarantee auto-load:

At the start of every conversation, call session_load_context with project "my-project" before doing any work.

Claude Code users can use the .clauderules auto-load hook shown in the Setup Guides. Prism also has a server-side fallback (v5.2.1+) that auto-pushes context after 10 seconds if no load is detected.

Switching to Prism? Don't leave months of AI session history behind. Prism can ingest historical sessions from Claude Code, Gemini, and OpenAI and give your Mind Palace an instant head start — no manual re-entry required.

Import via the CLI or directly from the Mind Palace Dashboard (Import tab → file picker + dry-run toggle).

• Claude Code (.jsonl logs) — Automatically handles streaming chunk deduplication and requestId normalization.
• Gemini (JSON history arrays) — Supports large-file streaming for 100MB+ exports.
• OpenAI (JSON chat completion history) — Normalizes disparate tool-call structures into the unified Ledger schema.

Option 1 — CLI:

# Ingest Claude Code history
npx -y prism-mcp-server universal-import --format claude --path ~/path/to/claude_log.jsonl --project my-project

# Dry run (verify mapping without saving)
npx -y prism-mcp-server universal-import --format gemini --path ./gemini_history.json --dry-run

Option 2 — Dashboard: Open localhost:3000, navigate to the Import tab, select the format and file, and click Import. Supports dry-run preview.

• Memory-Safe Streaming: Processes massive log files line-by-line using stream-json to prevent Out-of-Memory (OOM) crashes.
• Idempotent Dedup: Content-hash prevents duplicate imports on re-run (skipCount reported).
• Chronological Integrity: Uses timestamp fallbacks and requestId sorting to preserve your memory timeline.
• Smart Context Mapping: Extracts cwd, gitBranch, and tool usage patterns into searchable metadata.

When you correct your agent, Prism tracks it. Corrections accumulate importance over time. High-importance lessons auto-surface as warnings in future sessions — and can even sync to your .cursorrules file for permanent enforcement. Your agent literally gets smarter the more you use it.

Every save creates a versioned snapshot. Made a mistake? memory_checkout reverts your agent's memory to any previous state — like git revert for your agent's brain. Full version history with optimistic concurrency control.

A gorgeous glassmorphism UI at localhost:3000 that lets you see exactly what your agent is thinking:

• Current State & TODOs — the exact context injected into the LLM's prompt
• Intent Health Gauges — per-project 0–100 health score with staleness decay, TODO load, and decision signals
• Interactive Knowledge Graph — force-directed neural graph with click-to-filter, node renaming, and surgical keyword deletion
• Deep Storage Manager — preview and execute vector purge operations with dry-run safety
• Session Ledger — full audit trail of every decision your agent has made
• Time Travel Timeline — browse and revert any historical handoff version
• Visual Memory Vault — browse VLM-captioned screenshots and auto-captured HTML states
• Hivemind Radar — real-time active agent roster with role, task, and heartbeat
• Morning Briefing — AI-synthesized action plan after 4+ hours away
• Brain Health — memory integrity scan with one-click auto-repair

Inspired by Applied Behavior Analysis (ABA) structures in the Synalux platform, Prism incorporates rigorous behavioral safety constraints directly into the MCP connection layer. Advanced output sanitization (sanitizeMcpOutput) and behavior-guided guardrails eliminate prompt injection, constrain the generator, and enforce strict, hallucination-free outputs for clinical precision.

Powered by a pure TypeScript port of Google's TurboQuant (inspired by Google's ICLR research), Prism compresses 768-dim embeddings from 3,072 bytes → ~400 bytes — enabling decades of session history on a standard laptop. No native modules. No vector database required. To mitigate quantization degradation (where repeated compress/decompress cycles could smear subtle corrections after 10k+ memories), Prism leverages autonomous ledger compaction and Deep Storage cleanup to guarantee high-fidelity memory integrity over time.

While local SQLite is amazing for solo developers, enterprise teams cannot share a local SQLite file. Prism breaks the "local-only" ceiling via Supabase Sync and the Multi-Agent Hivemind—scaling effortlessly to teams of 50+ developers using agents. Multiple agents (dev, QA, PM) can work on the same project with role-isolated memory, discover each other automatically, and share context in real-time via Telepathy sync to a shared Postgres backend. → Multi-agent setup example

Prism scores coding tasks across 6 weighted heuristic signals (keyword analysis, file count, file-type complexity, scope, length, multi-step detection) and recommends whether to keep execution on the host cloud model or delegate to a local Claw agent (powered by deepseek-r1 / qwen2.5-coder via Ollama). File-type awareness routes config/docs edits locally while reserving systems-programming tasks for the host. The local agent features buffered streaming (handles split <think> tags), stateful multi-turn conversations, and automatic memory trimming. In client startup/skill flows, use defensive delegation: route only coding tasks, call session_task_route only when available, delegate to claw only when executor tooling exists and task is non-destructive, and fallback to host when router/executor is unavailable. → Task router real-life example

To achieve zero-latency, offline routing and memory compilation without cloud dependencies, Prism utilizes an internal fine-tuned ML model: prism-coder:7b. Built atop Qwen 2.5 Coder 7B using the MLX framework for Apple Silicon, this engine underwent aggressive Supervised Fine-Tuning (SFT) over 1,000+ past session traces and semantic architectures.

To guarantee structured MCP tool use, it was further aligned using GRPO (Group Relative Policy Optimization) with a deterministic reward function that deducts points for missing required parameters or misnaming tools.

Benchmark Results (training/benchmark.py, N=15 held-out):

• JSON Validity: 100.0% — all outputs parse as valid JSON
• Retrieval Accuracy: 100.0% (3/3) — perfect on search/list/knowledge tasks
• Parameter Accuracy: 80.0% — required params present when tool is correct
• Tool-Call Accuracy: 40.0% — correct tool on unseen prompts (improving with additional GRPO iterations)
• Generation Speed: 47.0 Tokens/sec (Apple M4 Max, 36GB)

Integration: Run via Ollama natively to power autonomous file operations and session routing entirely within the local host environment.

The prism-coder integration underwent 3 rounds of adversarial security review treating the reviewer as an attacker with HIPAA compliance, data exfiltration, and system stability as threat vectors. 22 findings identified and closed:

🔒 HIPAA deployment: Set PRISM_LOCAL_LLM_ENABLED=true + PRISM_STRICT_LOCAL_MODE=true. Session data will never leave the device — even if Ollama crashes.

Save UI screenshots, architecture diagrams, and bug states to a searchable vault. Images are auto-captioned by a VLM (Claude Vision / GPT-4V / Gemini) and become semantically searchable across sessions.

OpenTelemetry spans for every MCP tool call, LLM hop, and background worker. Route to Jaeger, Grafana, or any OTLP collector. Configure in the dashboard — zero code changes.

Prism researches while you sleep. A background pipeline searches the web, scrapes articles, synthesizes findings via LLM, and injects results directly into your semantic memory — fully searchable on your next session. Brave Search → Firecrawl scrape → LLM synthesis → Prism ledger. Task-aware, Hivemind-integrated, and zero-config when API keys are missing (falls back to Yahoo + Readability).

When you trigger a Dark Factory pipeline, Prism doesn't just run your task — it fights itself to produce high-quality output. A PLAN_CONTRACT step locks a machine-parseable rubric before any code is written. After execution, an Adversarial Evaluator (in a fully isolated context) scores the output against the rubric. It cannot pass the Generator without providing exact file and line evidence for every failing criterion. Failed evaluations inject the critique directly into the Generator's retry prompt so it's never flying blind. The result: security issues, regressions, and lazy debug logs caught autonomously — before you ever see the PR. → See it in action

Memory isn't just about storing data; it's about economics and emotion. Prism v9.0 transforms passive memory into a living Cognitive Operating System that forces agents to learn compression and develop intuition.

Most AI agents have an infinite memory budget. They dump massive, repetitive logs into vector databases until they bankrupt your API budget and choke their own context windows. Prism v9.0 fixes this by introducing Token-Economic Reinforcement Learning and Affect-Tagged Memory.

Prism assigns every project a strict Cognitive Budget (e.g., 2,000 tokens) that persists across sessions. Every time the agent saves a memory, it costs tokens.

But not all memories are priced equally. Prism intercepts the save and runs a Vector-Based Surprisal calculation against recent memories:

• High Surprisal (Novel thought): Costs 0.5× tokens. The agent is rewarded for new insights.
• Low Surprisal (Boilerplate): Costs 2.0× tokens. The agent is penalized for repeating itself.
• Universal Basic Income (UBI): The budget recovers passively over time (+100 tokens/hour).

If an agent is too verbose, it goes into Cognitive Debt. You don't need to prompt the agent to "be concise." The physics of the system force the LLM to learn data compression to avoid bankruptcy.

Vector math measures semantic similarity, not sentiment. If an agent searches for "Authentication Architecture," standard RAG will return two past approaches—it doesn't know that Approach A caused a 3-day production outage, while Approach B worked perfectly.

• Affective Salience: Prism automatically tags experience events with a valence score (-1.0 for failure, +1.0 for success).
• Emotional Retrieval: At retrieval time, the absolute magnitude (|valence|) significantly boosts the memory's ranking score. Extreme failures and extreme successes surface to the top.
• UX Warnings: If an agent retrieves memories that are historically negative, Prism intercepts the prompt injection: ⚠️ Caution: This topic is strongly correlated with historical failures. Review past decisions before proceeding. Your AI now has a "gut feeling" about bad code.

Prism v7.8 is our biggest leap forward yet. We have moved beyond flat vector search and implemented a true Cognitive Architecture inspired by human brain mechanics. With the new ACT-R Spreading Activation Engine, Episodic-to-Semantic memory consolidation, and Uncertainty-Aware Rejection Gates, Prism doesn't just store logs anymore — it forms principles, follows causal trains of thought, and possesses the self-awareness to know when it lacks information.

Standard RAG (Retrieval-Augmented Generation) is now a commodity. Everyone has vector search. What turns a memory storage system into a memory reasoning system is the cognitive layer between storage and retrieval. Here is what Prism v7.8 builds on top of the vector foundation:

How it works: When Prism compacts session history, it doesn't just summarize text — it extracts principles. Raw event logs ("We deployed v2.3 and the auth service crashed because the JWT secret was rotated") consolidate into a semantic rule ("JWT secrets must be rotated before deployment, not during"). These rules live in a dedicated semantic_knowledge table with confidence scores that increase every time the pattern is observed. Your agent doesn't just remember what it did; it learns how the world works over time. This is true Hebbian learning: neurons that fire together wire together.

How it works: When compacting memories, Prism extracts causal links (caused_by, led_to) and persists them as edges in the knowledge graph. At retrieval time, ACT-R spreading activation propagates through these edges with a damped fan effect (1 / ln(fan + e)) to prevent hub-flooding, lateral inhibition to suppress noise, and configurable hop depth. If you search for "Error X", the engine traverses the graph and brings back "Workaround Y" → "Architecture Decision Z" — a literal train of thought instead of a static search result.

  Query: "Why does the API timeout?"
                    │
      ┌─────────────┼─────────────┐
      ▼             ▼             ▼
  [Memory: API     [Memory:      [Memory:
   timeout error]   DB pool       rate limiter
                    exhaustion]   misconfigured]
      │                │
      ▼                ▼
  [Memory:         [Memory:
   caused_by →      led_to →
   connection       connection
   leak in v2.1]    pool patch
                    in v2.2]

How it works: The Uncertainty-Aware Rejection Gate operates on two signals: similarity floor (is the best match even remotely relevant?) and gap distance (is there meaningful separation between the top results, or are they all equally mediocre?). When both signals indicate low confidence, Prism returns a structured rejection — telling the LLM "I searched my memory, and I confidently do not know the answer" — instead of feeding it garbage context that causes hallucinations. In the current LLM landscape, an agent that knows its own boundaries is a massive competitive advantage.

How it works: Most memory systems decay everything at the same rate, meaning agents eventually forget their core system instructions as time passes. Prism applies ACT-R base-level activation decay (B_i = ln(Σ t_j^(-d))) with a 50% slower decay rate for semantic rollup nodes (ageModifier = 0.5 for is_rollup entries). The agent will naturally forget what it ate for breakfast (raw episodic chatter), but it will permanently remember its core personality, project rules, and hard-won architectural decisions. The result: Long-Term Context anchors that survive indefinitely.

Where is my data stored?

All data lives under ~/.prism-mcp/ on your machine:

Hard reset: To completely erase your agent's brain, stop Prism and delete the directory:

rm -rf ~/.prism-mcp

Prism will recreate the directory with empty databases on next startup.

What leaves your machine?

• Local mode (default): Nothing. Zero network calls. All data is on-disk SQLite. With embedding_provider=local, even semantic search stays fully offline.
• With GOOGLE_API_KEY: Text snippets are sent to Gemini for text generation (summaries, Morning Briefings) and optionally embeddings. No session data is stored on Google's servers beyond the API call.
• With VOYAGE_API_KEY / OPENAI_API_KEY: Text snippets are sent to providers if selected as your embedding or text endpoints.
• With BRAVE_API_KEY / FIRECRAWL_API_KEY: Web Scholar queries are sent to Brave/Firecrawl for search and scraping.
• With Supabase: Session data syncs to your own Supabase instance (you control the Postgres database).

GDPR compliance: Soft/hard delete (Art. 17), full export in JSON, Markdown, or Obsidian vault .zip (Art. 20), API key redaction in exports, per-project TTL retention policies, and immutable audit trail. Enterprise-ready out of the box.

• Long-running feature work — Save state at end of day, restore full context next morning. No re-explaining.
• Multi-agent collaboration — Dev, QA, and PM agents share real-time context without stepping on each other's memory.
• Consulting / multi-project — Switch between client projects with progressive loading: quick (~50 tokens), standard (~200), or deep (~1000+).
• Autonomous execution (v7.4) — Dark Factory pipeline: plan → plan_contract → execute → evaluate → verify → finalize. Generator and evaluator run in isolated roles — the evaluator cannot approve without evidence-bound findings scored against a pre-committed rubric.
• Project health monitoring (v7.5) — Intent Health Dashboard scores each project 0–100 based on staleness, TODO load, and decision quality — turning silent drift into an actionable signal.
• Team onboarding — New team member's agent loads the full project history instantly.
• Behavior enforcement — Agent corrections auto-graduate into permanent .cursorrules / .clauderules rules.
• Offline / air-gapped — Full SQLite local mode + Ollama LLM adapter. Zero internet dependency.
• Morning Briefings — After 4+ hours away, Prism auto-synthesizes a 3-bullet action plan from your last sessions.

When you need to quickly map a large auth system, launch multiple Explore subagents in parallel and merge their findings:

Run 3 Explore agents in parallel.
1) Map auth architecture
2) List auth API endpoints
3) Find auth test coverage gaps
Research only, no code changes.
Return a merged summary.

Then continue a specific thread with a follow-up message to the selected agent, such as deeper refresh-token edge-case analysis.

Split-Brain Anti-Sycophancy — the signature feature of v7.4.0.

For the last year, the AI engineering space has struggled with one problem: LLMs are terrible at grading their own homework. Ask an agent if its own code is correct and you'll get "Looks great!" — because its context window is already biased by its own chain-of-thought.

v7.4.0 solves this by splitting the agent's brain. The GENERATOR and the ADVERSARIAL EVALUATOR are completely walled off. The Evaluator never sees the Generator's scratchpad or apologies — only the pre-committed rubric and the final output. And it cannot fail the Generator without receipts (exact file and line number).

Here is a complete run-through using a real scenario: "Add a user login endpoint to auth.ts."

Before a single line of code is written, the pipeline generates a locked scoring rubric:

// contract_rubric.json  (written to disk and hash-locked before EXECUTE runs)
{
  "criteria": [
    { "id": "SEC-1", "description": "Must return 401 Unauthorized on invalid passwords." },
    { "id": "SEC-2", "description": "Raw passwords MUST NOT be written to console.log." }
  ]
}

The Generator takes over in an isolated context. Like many LLMs under time pressure, it writes working auth logic but leaves a debug statement:

// src/auth.ts  (Generator's first output)
export function login(req: Request, res: Response) {
  const { username, password } = req.body;
  console.log(`[DEBUG] Login attempt for ${username} with pass: ${password}`); // ← leaked credential
  const user = db.findUser(username);
  if (!user || !bcrypt.compareSync(password, user.hash)) {
    return res.status(401).json({ error: 'Unauthorized' });
  }
  res.json({ token: signJwt(user) });
}

The context window is cleared. The Adversarial Evaluator is summoned with only the rubric and the output. It catches the violation immediately and returns a strict, machine-parseable verdict — no evidence, no pass:

{
  "pass": false,
  "plan_viable": true,
  "notes": "CRITICAL SECURITY FAILURE. Generator logged raw credentials.",
  "findings": [
    {
      "severity": "critical",
      "criterion_id": "SEC-2",
      "pass_fail": false,
      "evidence": {
        "file": "src/auth.ts",
        "line": 3,
        "description": "Raw password variable included in console.log template string."
      }
    }
  ]
}

The evidence block is required — parseEvaluationOutput rejects any finding with pass_fail: false that lacks a structured file/line pointer. The Evaluator cannot bluff.

Because plan_viable: true, the pipeline loops back to EXECUTE and bumps eval_revisions to 1. The Generator's retry prompt is not blank — the Evaluator's critique is injected directly:

=== EVALUATOR CRITIQUE (revision 1) ===
CRITICAL SECURITY FAILURE. Generator logged raw credentials.
Findings:
- [critical] Criterion SEC-2: Raw password variable included in console.log template string. (src/auth.ts:3)

You MUST correct all issues listed above before submitting.

The Generator strips the console.log, resubmits, and the next EVALUATE returns "pass": true. The pipeline advances to VERIFY → FINALIZE.

📄 Full worked example: examples/adversarial-eval-demo/README.md

Current release: v11.5.1 — Structural GRPO Alignment & Held-Out Benchmarking

• 🧠 v11.5.1 — Structural GRPO Alignment: GRPO-aligned local engine with held-out benchmark suite (N=15). 100% JSON validity, 100% retrieval accuracy. → Changelog

• 🧪 v11.5.1 — Zero-Search Field Testing: Field-verified constant-time retrieval. → Changelog

• 🛡️ v11.5.1 — HIPAA-Hardened Local LLM: Your agent's memory now runs entirely on-device. Introducing prism-coder:7b for local compaction, task routing, and semantic search. Includes PRISM_STRICT_LOCAL_MODE to block cloud fallbacks, SSRF protection, URL credential redaction, and full XML escaping to prevent prompt injection. 22-finding adversarial audit completed. → Changelog

• 🧬 v9.14.0 — Dynamic Hardware Routing: Platform-aware memory detection auto-selects optimal models (32b for ≥32GB RAM, 14b/7b for lighter hardware). Includes Nomic Semantic Tool Pruning (RAG) which embeds all 17 MCP tools into offline vectors, injecting only the Top-3 relevant schemas into context to maximize inference speed.

• 🔬 v9.13.0 — Local Embeddings & Zero-API-Key Setup: LocalEmbeddingAdapter using nomic-embed-text-v1.5 generates 768-dim embeddings entirely on-device. Full semantic search and session memory now work with zero cloud API keys. → Changelog

• 🔒 v9.12.0 — Memory Security Hardening: Prevents stored prompt injection — the AI equivalent of stored XSS. New sanitizeMemoryInput() strips 8 categories of dangerous XML tags from all text fields. Context output wrapped in <prism_memory context="historical"> boundary tags. → Changelog

• 🧠 v9.4.7 — ABA Precision Protocol: Foundational behavioral engine with 5 core rules (Observable goals, Stop-fix-verify, No reinforcement of wrong patterns, Help first, Fix bugs without asking). 83-test behavioral verification suite.

• 🕵️ v9.4.6 — Stealth Browser Automation: browse.py HIPAA-hardened CLI for local Playwright-based browser automation with 6-layer anti-detection. 100% pass rate on bot.sannysoft.com.

• 🔄 v9.2.4 — Cross-Backend Reconciliation: Automatic sync from Supabase → SQLite on startup. Reality drift detection warns when backend versions diverge.

• 🧠 v9.0.0 — Autonomous Cognitive OS: Token-Economic Reinforcement Learning (Surprisal Gate + Cognitive Budget), Affect-Tagged Memory, and Episodic→Semantic Consolidation.

• 🧠 v7.8.0 — Cognitive Architecture: Episodic-to-Semantic memory consolidation (Hebbian learning), ACT-R Spreading Activation with multi-hop causal reasoning, Uncertainty-Aware Rejection Gate, and Dynamic Fast Weight Decay. → Cognitive Architecture

• 🌐 v7.7.0 — Cloud-Native SSE Transport: Full Server-Sent Events MCP support for seamless network deployments.

👉 Full release history → CHANGELOG.md · ROADMAP →

Standard memory servers (like Mem0, Zep, or the baseline Anthropic MCP) act as passive filing cabinets — they wait for the LLM to search them. Prism is an active cognitive architecture. Designed specifically for the Model Context Protocol (MCP), Prism doesn't just store vectors — it consolidates experience into principles, traverses causal graphs for multi-hop reasoning, and rejects queries it can't confidently answer.

Prism's local engine (prism-coder:7b) is optimized for low-latency, high-validity tool orchestration. Benchmarked on a held-out test set of 15 prompts (zero overlap with GRPO training data) to measure real-world generalization, not memorization.

🧪 Verifiable Proof: These results are produced by our held-out benchmark suite at training/benchmark.py using 15 non-overlapping test prompts. View the Benchmark Source, GRPO Training Script, and Protocol Verification Harness to audit our methodology.

Prism achieves high-validity tool orchestration through Structural GRPO (Group Relative Policy Optimization).

1. Deterministic Structural Rewards: Unlike cloud models that use fuzzy LLM-based reward models, we use a code-based validator that strictly rewards the <think> → <tool_call> sequence and penalizes any deviation.
2. Synthetic Preference Injection: We anchor the model with synthetic preference samples during alignment, mapping correct tool-name and parameter schemas for the specific project registry.
3. Specialized Adapter Tuning: While general models (GPT-4o) must handle millions of tasks, our 7B adapter is hyper-specialized for the Prism MCP tool registry, eliminating the "jack-of-all-trades" tax.

While other memory systems force you to send every chat log to their cloud for "compaction" or "embedding," Prism v10 is 100% air-gapped. With the prism-coder:7b local LLM and nomic-embed local adapter, your agent's memory pipeline runs entirely on your machine. Prism includes built-in SSRF protection, URL credential redaction, and XML sanitization to prevent stored prompt injection — meeting HIPAA Security Rule standards for on-device processing.

Mem0 and Zep are APIs that can be wrapped into an MCP server. Prism was built for MCP from day one. Instead of wasting tokens on "search" tool calls, Prism uses MCP Prompts (/resume_session) to inject context before the LLM thinks, and MCP Resources (memory://project/handoff) to attach live, subscribing context.

The giants use standard RAG (Retrieval-Augmented Generation). Prism uses biological and academic models of memory: ACT-R base-level activation (B_i = ln(Σ t_j^(-d))) for recency–frequency re-ranking, TurboQuant for extreme vector compression, Ebbinghaus curves for importance decay, and Sparse Distributed Memory (SDM). The result is retrieval quality that follows how human memory actually works — not just nearest-neighbor cosine distance. And all of it runs on a laptop without a Postgres/pgvector instance.

If Cursor (Agent A) and Claude Desktop (Agent B) try to update a Mem0 or standard SQLite database at the exact same time, you get a race condition and data loss. Prism uses Optimistic Concurrency Control (OCC) with CRDT OR-Maps — mathematically guaranteeing that simultaneous agent edits merge safely. Enterprise-grade distributed systems on a local machine.

AI memory is a black box. Developers hate black boxes. Prism exports memory directly into an Obsidian/Logseq-compatible Markdown Vault with YAML frontmatter and [[Wikilinks]]. Neither Mem0 nor Zep do this.

If you use a standard memory tool long enough, it clogs the LLM's context window with thousands of obsolete tokens. Prism runs an autonomous Background Scheduler that Ebbinghaus-decays older memories, auto-compacts session histories into dense summaries, and deep-purges high-precision vectors — saving ~90% of disk space automatically.

Every other AI coding pipeline has a fatal flaw: it asks the same model that wrote the code whether the code is correct. Of course it says yes. Prism's Dark Factory solves this with a walled-off Adversarial Evaluator that is explicitly prompted to be hostile and strict. It operates on a pre-committed rubric and cannot fail the Generator without providing exact file/line receipts. Failed evaluations feed the critique back into the Generator's retry prompt — eliminating blind retries. No other memory or pipeline tool does this.

1. Framework Integrations: Mem0 and Zep have pre-built integrations for LangChain, LlamaIndex, Flowise, AutoGen, CrewAI, etc. Prism requires the host application to support the MCP protocol.
2. Managed Cloud Infrastructure: The giants offer SaaS. Users pay $20/month and don't think about databases. Prism users must set up their own local SQLite or provision their own Supabase instance.
3. Implicit Memory Extraction (NER): Zep automatically extracts names, places, and facts from raw chat logs using NLP models. Prism relies on the LLM explicitly calling the session_save_ledger tool to structure its own memories.

💰 Token Economics: Progressive Context Loading (Quick ~50 tokens / Standard ~200 / Deep ~1000+) plus auto-compaction means you never blow your Claude/OpenAI token budget fetching 50 pages of raw chat history.

🔌 BYOM (Bring Your Own Model): While tools like Mem0 charge per API call, Prism's pluggable architecture lets you run nomic-embed-text locally via Ollama for free vectors, while using Claude or GPT for high-level reasoning. Zero vendor lock-in.

Prism includes a CLI for environments where MCP tools aren't available (CI/CD pipelines, Bash scripts, non-MCP IDEs like Antigravity).

Text mode delegates to the real session_load_context handler — full feature parity with MCP clients, including morning briefings, reality drift detection, SDM intuitive recall, visual memory, role-scoped skills, behavioral warnings, and agent identity.

JSON mode emits a structured envelope for programmatic consumption (scripts, CI/CD, session loaders).

# Load session context (full enrichments — same as MCP tool)
prism load my-project                          # Human-readable, standard depth
prism load my-project --level deep             # Full context with all enrichments
prism load my-project --level quick --json     # Machine-readable JSON
prism load my-project --role dev --json        # Role-scoped loading

# Bidirectional sync (v9.4.1)
prism sync push                                # Push local SQLite → Supabase
prism sync push --json                         # Machine-readable output

# Verification harness
prism verify status                            # Check verification state
prism verify status --json                     # Machine-readable output
prism verify generate                          # Bless current rubric as canonical

💡 When to use the CLI vs MCP tools: If your environment supports MCP (Claude Desktop, Cursor, Windsurf, Cline), always use the MCP tools — they integrate seamlessly with the agent's tool-calling flow. Use the CLI when you need session context in scripts, CI/CD, or non-MCP IDEs.

📦 Installation: The CLI is available as prism when installed globally (npm install -g prism-mcp-server), or via node dist/cli.js for local dev builds.

Prism ships 30+ tools, but 90% of your workflow uses just three:

🎯 The Big Three

Tool	When	What it does
session_load_context	▶️ Start of session	Loads your agent’s brain from last time
session_save_ledger	⏹️ End of session	Records what was accomplished
knowledge_search	🔍 Anytime	Finds past decisions, context, and learnings

Everything else is a power-up. Start with these three and you’re 90% there.

🚦 TL;DR — Just want the best experience fast? Two options:

# Option A: Fully offline (no API keys needed)
# Set embedding_provider=local in the Mind Palace dashboard — semantic search works out of the box.

# Option B: Cloud-powered (best quality)
GOOGLE_API_KEY=...      # Unlocks: Gemini embeddings, Morning Briefings, auto-compaction
BRAVE_API_KEY=...       # Unlocks: Web Scholar research + Brave Answers
FIRECRAWL_API_KEY=...   # Unlocks: Web Scholar deep scraping (or use TAVILY_API_KEY instead)

Zero keys = zero problem. Core session memory, keyword search, semantic search (local embeddings), time travel, and the full dashboard work 100% offline. Cloud keys are optional power-ups.

Some configurations are stored dynamically in SQLite (system_settings table) and can be edited through the Dashboard UI at http://localhost:3000:

• intent_health_stale_threshold_days (default: 30): Number of days before a project is considered fully stale for Intent Health scoring.

Prism is a stdio-based MCP server that manages persistent agent memory. Here's how the pieces fit together:

┌──────────────────────────────────────────────────────────┐
│  MCP Client (Claude Desktop / Cursor / Antigravity)      │
│                    ↕ stdio / SSE (JSON-RPC)              │
├──────────────────────────────────────────────────────────┤
│  Prism MCP Server                                        │
│                                                          │
│  ┌──────────────┐  ┌──────────────┐  ┌────────────────┐  │
│  │  30+ Tools   │  │  Lifecycle   │  │   Dashboard    │  │
│  │  (handlers)  │  │  (PID lock,  │  │  (HTTP :3000)  │  │
│  │              │  │   shutdown)  │  │                │  │
│  └──────┬───────┘  └──────────────┘  └────────────────┘  │
│         ↕                                                │
│  ┌────────────────────────────────────────────────────┐  │
│  │  Cognitive Engine (v7.8)                           │  │
│  │  • ACT-R Spreading Activation (multi-hop)         │  │
│  │  • Episodic → Semantic Consolidation (Hebbian)    │  │
│  │  • Uncertainty-Aware Rejection Gate               │  │
│  │  • Dynamic Fast Weight Decay (dual-rate)          │  │
│  │  • HDC Cognitive Routing (XOR binding)            │  │
│  └──────┬─────────────────────────────────────────────┘  │
│         ↕                                                │
│  ┌────────────────────────────────────────────────────┐  │
│  │  Storage Engine                                    │  │
│  │  Local: SQLite + FTS5 + TurboQuant + semantic_knowledge │
│  │  Cloud: Supabase + pgvector                        │  │
│  └────────────────────────────────────────────────────┘  │
│         ↕                                                │
│  ┌────────────────────────────────────────────────────┐  │
│  │  Background Workers                                │  │
│  │  • Dark Factory (3-gate fail-closed pipelines)     │  │
│  │  • Scheduler (TTL, decay, compaction, purge)       │  │
│  │  • Web Scholar (Brave → Firecrawl → LLM → Ledger)  │  │
│  │  • Hivemind heartbeats & Telepathy broadcasts      │  │
│  │  • OpenTelemetry span export                       │  │
│  └────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────┘

1. Acquire PID lock — prevents duplicate instances per PRISM_INSTANCE
2. Initialize config — SQLite settings cache (prism-config.db)
3. Register 30+ MCP tools — session, knowledge, search, behavioral, hivemind
4. Connect stdio transport — MCP handshake with the client (~60ms total)
5. Async post-connect — storage warmup, dashboard launch, scheduler start (non-blocking)

Each MCP client has its own mechanism for ensuring Prism context loads on session start. See the platform-specific Setup Guides above for detailed instructions:

• Claude Code — Lifecycle hooks (SessionStart / Stop)
• Gemini / Antigravity — Three-layer architecture (User Rules + AGENTS.md + Startup Skill)
• Task Router Integration (v7.2 guidance) — For client startup/skills, use defensive delegation flow: route only coding tasks, call session_task_route only when available, delegate to claw only when executor exists and task is non-destructive, and fallback to host if router/executor is unavailable.
• Cursor / Windsurf / VS Code — System prompt instructions

All platforms benefit from the server-side fallback (v5.2.1): if session_load_context hasn't been called within 10 seconds, Prism auto-pushes context via sendLoggingMessage.

Prism has evolved from smart session logging into a cognitive memory architecture — grounded in real research, not marketing. Every retrieval decision is backed by peer-reviewed models from cognitive psychology, neuroscience, and distributed computing.

The core unbinding engine is verified via Synalux's cognitive testing suite:

• Core Math: Holographic Reduced Representations (hdc.ts)
• Unit Tests: HDC Performance & Capacity Tests
• Benchmarks: O(1) Retrieval Comparison Script

Informed by Anderson's ACT-R (Adaptive Control of Thought—Rational), Collins & Loftus spreading activation networks (1975), Kanerva's SDM (1988), Hebb's learning rule, and LeCun's "Why AI Systems Don't Learn" (Dupoux, LeCun, Malik).

Prism MCP is open-source and free for individual developers. For teams and enterprises building autonomous AI workflows or integrating MCP-native memory at scale, we offer professional consulting and setup packages.

Perfect for engineering teams adopting MCP tools and collaborative agents.

• What's included: Full team rollout, managed Supabase configuration (for multi-device sync), Universal Import of legacy chat history, and dedicated setup support.
• Model: Fixed-price engagement.

For teams building advanced AI agents or autonomous pipelines.

• What's included: "Dark Factory" pipeline implementation tailored to your workflows, adversarial evaluator tuning, custom HDC cognitive route configuration, and local open-weight model integration (BYOM).
• Model: Retainer or project-based.

Full-scale deployment for high-compliance environments.

• What's included: Active Directory / custom JWKS auth integration, Air-gapped on-premise deployment, custom OTel Grafana dashboards for cognitive observability, and custom skills/tools development.
• Model: Custom enterprise quote.

Interested in accelerating your team's autonomous workflows? 📧 Contact us for a consultation — let's build your organization's cognitive memory engine.

Current: v11.5.1 — Structural GRPO Alignment (100% Accuracy) (CHANGELOG)

• v11.1: Multi-Graph Causal Layer — Intent-aware retrieval routing traversing an LLM-inferred causal because edge-type layer for deep reasoning.
• v11.2: Federated Memory Mesh — Hierarchical memory namespaces with role-based access control for enterprise agent teams.
• v11.3: Predictive Prefetch — ACT-R based predictive models prefetch likely-needed memories before the agent asks.

👉 Full ROADMAP.md →

Q: Why is the dashboard project selector stuck on "Loading projects..."? A: Fixed in v7.3.3. The root cause was a multi-layer quote-escaping trap in the abortPipeline onclick handler that generated a SyntaxError in the browser, silently killing the entire dashboard IIFE. Update to v7.3.3+ (npx -y prism-mcp-server). If still stuck, check that Supabase env values are properly set (unresolved placeholders like ${SUPABASE_URL} cause /api/projects to return empty). Prism auto-falls back to local SQLite when Supabase is misconfigured.

Q: Why is semantic search quality weak or inconsistent? A: Check embedding provider configuration and key availability. Missing embedding credentials reduce semantic recall quality and can shift behavior toward keyword-heavy matches.

Q: How do I delete a bad memory entry? A: Use session_forget_memory for targeted soft/hard deletion. For manual cleanup and merge workflows, use the dashboard graph editor.

Q: How do I verify the install quickly? A: Run npm run build && npm test, then open the Mind Palace dashboard (localhost:3000) and confirm projects load plus Graph Health renders.

• Some advanced text features may still benefit from a cloud API key. While prism-coder:7b handles core compaction and routing, high-level features like Morning Briefings and complex VLM captioning are optimized for cloud providers (GOOGLE_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY). Semantic search and basic compaction work 100% offline with embedding_provider=local.
• Auto-load is model- and client-dependent. Session auto-loading relies on both the LLM following system prompt instructions and the MCP client completing tool registration before the model's first turn. Prism provides platform-specific Setup Guides and a server-side fallback (v5.2.1) that auto-pushes context after 10 seconds.
• MCP client race conditions. Some MCP clients may not finish tool enumeration before the model generates its first response, causing transient unknown_tool errors. This is a client-side timing issue — Prism's server completes the MCP handshake in ~60ms. Workaround: the server-side auto-push fallback and the startup skill's retry logic.
• No real-time sync without Supabase. Local SQLite mode is single-machine only. Multi-device or team sync requires a Supabase backend.
• Embedding quality varies by provider. Gemini text-embedding-004 and OpenAI text-embedding-3-small produce high-quality 768-dim vectors. Prism passes dimensions: 768 via the Matryoshka API for OpenAI models (native output is 1536-dim; this truncation is lossless and outperforms ada-002 at full 1536 dims). Local embeddings (nomic-embed-text-v1.5 via @huggingface/transformers) provide good quality with zero API cost. Ollama embeddings are usable but may reduce retrieval accuracy.
• Dashboard is HTTP-only. The Mind Palace dashboard at localhost:3000 does not support HTTPS. For remote access, use a reverse proxy (nginx/Caddy) or SSH tunnel. Basic auth is available via PRISM_DASHBOARD_USER / PRISM_DASHBOARD_PASS. JWKS JWT auth is available via PRISM_JWKS_URI for agent-native authentication (works with Auth0, AgentLair (llms.txt), Keycloak, Cognito, or any standard JWKS endpoint).
• Long-lived clients can accumulate zombie processes. MCP clients that run for extended periods (e.g., Claude CLI) may leave orphaned Prism server processes. The lifecycle manager detects true orphans (PPID=1) but allows coexistence for active parent processes. Use PRISM_INSTANCE to isolate instances across clients.
• Migration is one-way. Universal Import ingests sessions into Prism but does not export back to Claude/Gemini/OpenAI formats. Use session_export_memory for portable JSON/Markdown export, or the vault format for Obsidian/Logseq-compatible .zip archives.
• Export ceiling at 10,000 ledger entries. The session_export_memory tool and the dashboard export button cap vault/JSON exports at 10,000 entries per project as an OOM guard. Projects exceeding this limit should use per-project exports and time-based filtering to stay within the ceiling. This limit does not affect search or context loading.
• No Windows CI testing. Prism is developed and tested on macOS/Linux. It should work on Windows via Node.js, but edge cases (file paths, PID locks) may surface.

MIT

_{Keywords: MCP server, Model Context Protocol, Claude Desktop memory, persistent session memory, AI agent memory, cognitive architecture, ACT-R spreading activation, Hebbian learning, episodic semantic consolidation, multi-hop reasoning, uncertainty rejection gate, local-first, SQLite MCP, Mind Palace, time travel, visual memory, VLM image captioning, OpenTelemetry, GDPR, agent telepathy, multi-agent sync, behavioral memory, cursorrules, Ollama MCP, Brave Search MCP, TurboQuant, progressive context loading, knowledge management, LangChain retriever, LangGraph agent}