# rasputin-memory

> The memory system your AI agent deserves. 4-stage hybrid retrieval — Vector + BM25 + Knowledge Graph + Neural Reranker — in <150ms. Self-hosted, $0/query, built for agents that need to actually rememb

- **URL**: https://www.freshcrate.ai/projects/rasputin-memory
- **Author**: jcartu
- **Category**: Databases
- **Latest version**: `v0.9.1` (2026-04-16)
- **License**: MIT
- **Source**: https://github.com/jcartu/rasputin-memory
- **Language**: Python
- **GitHub**: 30 stars, 5 forks
- **Registry**: github
- **Tags**: `agent-memory`, `ai`, `ai-memory`, `bm25`, `embeddings`, `falkordb`, `hybrid-search`, `inference`, `python`, `rag`

## Description

The memory system your AI agent deserves. 4-stage hybrid retrieval — Vector + BM25 + Knowledge Graph + Neural Reranker — in <150ms. Self-hosted, $0/query, built for agents that need to actually remember.

## Recent releases

| Version | Date | Urgency | Changes |
| --- | --- | --- | --- |
| `v0.9.1` | 2026-04-16 | High | ## What's New in v0.9.1  ### Semantic kNN Graph Expansion (Experimental) - Gated behind `KNN_LINKS=1` (off by default) - At ingest: each fact linked to top-30 similar existing facts (cosine >= 0.6) via Qdrant payload `similar_ids` - At search: fact-lane seeds expanded through links before CE reranking (capped at 10 expansions) - Architectural parity with Hindsight's `link_expansion_retrieval.py` - Full 10-conv benchmark: 72.1% non-adv (−2.1pp from baseline) — useful for graph-traversal workloads |
| `v0.9.0` | 2026-04-13 | High | ## Qwen3-Reranker + BM25 FTS5 + Prompt Routing  **Production: 74.2% non-adv** (+6.7pp from baseline). **Compare: 77.7% non-adv** (+10.2pp from baseline).  Full 10-conversation LoCoMo evaluation (1986 questions). 30+ documented experiments.  ### Benchmark Results  \| Category \| Production \| Compare \| Questions \| Δ from v0.8 \| \|----------\|-----------\|---------\|-----------\|-------------\| \| **Overall non-adv** \| **74.2%** \| **77.7%** \| 1540 \| +5.1pp / +8.6pp \| \| Open-domain \| 84.8% \| 83.2% \| 841 \| +3 |
| `v0.8.0` | 2026-04-10 | High | ## Full 10-Conversation LoCoMo Validation  **69.1% non-adversarial** (1986 questions, production mode). 21 documented experiments with scientific methodology.  ### Benchmark Results (LoCoMo full 10-conv, production mode)  \| Category \| Accuracy \| Questions \| Notes \| \|----------\|----------\|-----------\|-------\| \| Open-domain \| 81.1% \| 841 \| Rock solid \| \| Temporal \| 66.4% \| 321 \| 61% of failures are generation, not retrieval \| \| Multi-hop \| 55.2% \| 96 \| +16.7pp from prompt routing \| \| Single-hop \| |
| `v0.7.0` | 2026-04-03 | High | ## #1 on LoCoMo — 91.36%  Three benchmarks, one pipeline. All results are reproducible from the scripts in `benchmarks/`.  ### Benchmark Results  \| Benchmark \| Score \| Questions \| Venue \| \|-----------\|-------\|-----------\|-------\| \| **LoCoMo** \| **91.36%** (#1) \| 1,986 \| ACL 2024 \| \| **LongMemEval** \| **89.40%** \| 500 \| ICLR 2025 \| \| **FRAMES** \| **50.4%** \| 824 \| Google 2024 \|  #### LoCoMo Leaderboard  \| Rank \| System \| Accuracy \| \|------\|--------\|----------\| \| **#1** \| **RASPUTIN Memory v0.7** |
| `v0.6.0` | 2026-04-02 | Medium | ## RASPUTIN Memory v0.6.0 — #2 on LoCoMo (89.81%)  ### LoCoMo Benchmark Results  \| Rank \| System \| Accuracy \| \|------\|--------\|----------\| \| 🥇 \| Backboard \| 90.00% \| \| 🥈 \| **RASPUTIN** \| **89.81%** \| \| 🥉 \| Memvid \| 85.70% \| \| 4 \| MemMachine \| 84.87% \| \| 5 \| Memobase \| 75.78% \| \| 6 \| Zep \| 75.14% \| \| 7 \| mem0 \| 66.88% \|  **Config:** nomic-embed-text (768d) → Qdrant top-60 → Claude Opus (answer gen) → GPT-4o-mini (judge)  Full changelog: https://github.com/jcartu/rasputin-memory/blob/main/CHANG |
| `v0.5.0` | 2026-04-01 | Medium | ## Search Quality Breakthrough  Keyword overlap boosting + entity-aware scoring push recall well past mem0 benchmarks.  ### Benchmarks  \| Metric \| v0.4.0 \| v0.5.0 \| Change \| \|--------\|--------\|--------\|--------\| \| recall@5 \| 0.67 \| **0.82** \| +22% \| \| recall@10 \| 0.745 \| **0.885** \| +19% \| \| MRR@10 \| 0.56 \| **0.68** \| +21% \| \| Entity recall@5 \| 0.20 \| **0.63** \| +215% \| \| Decay recall@5 \| 0.23 \| **0.40** \| +74% \| \| Contradiction recall@5 \| 0.48 \| **0.96** \| +100% \|  Beats mem0 LOCOMO benchmark ( |
| `v0.4.0` | 2026-04-01 | Medium | ## What's Changed  ### Architecture - **Modular codebase**: 1,800-line `hybrid_brain.py` split into 11 focused modules in `brain/` package - **Unified scoring**: 5 competing source-weight systems replaced by single `scoring_constants.py` - **Shared utilities**: Extracted locking, batch Qdrant scroll, date parsing — eliminated 4x copy-paste patterns  ### Retrieval Quality - **Language-agnostic pipeline**: Deleted English-only keyword routing, stop words, and supersedes token checks — the embeddin |
| `v0.3.0` | 2026-03-30 | Medium | ## What's in v0.3.0  ### BEIR Benchmarks Real retrieval evaluation on BEIR datasets (SciFact, NFCorpus) using the full local infrastructure: - **SciFact**: Hybrid NDCG@10=0.8336 vs Vector-only 0.8230 (+0.011) - **NFCorpus**: Vector baseline NDCG@10=0.371 (strong MoE embeddings) - Full reproduction script: `benchmarks/run_beir.py`  ### Ablation Study 4-stage pipeline contribution analysis: - RRF fusion boosts Recall@10 by +1.5% (SciFact) - Neural reranker improves MRR@10 by +2.3% (SciFact)   - Gr |
| `v0.2.0` | 2026-03-30 | Medium | ## What changed  ### PII Scrub - Removed all personal data from code, docs, and examples (health metrics, personal names, specific locations, crypto references) - Replaced hardcoded `~/.openclaw/workspace` paths with environment variables (`WORKSPACE_PATH`, `ENTITY_GRAPH_PATH`, etc.) - Generalized example data in docstrings and documentation  ### Architecture Fixes - **Removed expansion maps** from `memory_engine.py`: the hand-rolled keyword topic tables were architectural debt compensating for |
| `v0.1.0` | 2026-03-30 | Medium | Initial public release of the RASPUTIN Memory System.  ## Features - 4-stage hybrid retrieval pipeline (Vector + BM25 + Graph + Reranker) - Multi-tenant agent isolation - Unified consolidation engine - A-MAC quality gate for memory commits - Docker Compose deployment - Dockerfile for standalone brain server - Predictive memory prefetch - STORM wiki generation from memory |

## Citation

- HTML: https://www.freshcrate.ai/projects/rasputin-memory
- Markdown: https://www.freshcrate.ai/projects/rasputin-memory.md
- Dependencies JSON: https://www.freshcrate.ai/api/projects/rasputin-memory/deps

_Generated by freshcrate.ai. Indexes github releases for AI-agent ecosystem packages._