内容
# memoirs
> **memoirs** /ˈmɛmwɑrz/ · _MEM-wahrz_ · from French *mémoires* (memories) — the durable written record an agent keeps of what it has learned.
**Local-first long-term memory for AI agents.**
[](./tests)
[](./scripts/coverage.sh)
[](./pyproject.toml)
[](./LICENSE)
memoirs gives every AI agent on your machine a single, durable memory layer that survives across sessions, IDEs, and even models. It ingests your transcripts, extracts the durable signal (preferences, decisions, projects, tool-calls), curates them with a local LLM, and serves the right ~1K-token context the moment your agent asks.
```python
# In your agent
context = mcp.call("mcp_get_context", {"query": "what's our auth strategy?"})
# → ranked memories, conflict-resolved, in 6.2 ms p50
```
No cloud. No API keys leaving the box. SQLite + a 2 GB GGUF model — it runs on a laptop.
---
## Why memoirs
| | memoirs |
|---|---|
| **100% local** | No data egress. No SaaS. The DB is one file you can `cp`, `rsync`, or encrypt with SQLCipher. |
| **Universal ingest** | Claude Code transcripts, Cursor `state.vscdb`, ChatGPT zip, Claude.ai export, JSONL, Markdown. One command. |
| **Native MCP** | 22 tools served over stdio. Drop into Claude Desktop, Claude Code, Cursor, Continue, Cline — pick one or all. |
| **Hybrid retrieval** | BM25 + dense + RRF, with optional graph multi-hop (HippoRAG-style PPR) and RAPTOR hierarchical summaries. |
| **Temporal & explainable** | Bi-temporal validity, `as_of=t` time-travel queries, full provenance chain (memory ← candidate ← message ← conversation ← source). |
| **LLM curator built-in** | Qwen 2.5 3B local for extraction, consolidation, conflict resolution. Auto-detects available models, falls back to heuristics. |
| **Auto-prune & forget** | Ebbinghaus decay, sleep-time consolidation cron, 4-mode Zettelkasten linking, EXPIRE/ARCHIVE generation. |
| **Privacy-first** | PII redaction (Presidio + secret scanning), per-memory ACL with visibility tiers, GDPR export/import portable bundles, encryption-at-rest (SQLCipher). |
---
## Install
memoirs ships three install paths. Pick one.
### 1. PyPI (recommended for users)
```bash
# minimal: SQLite + FTS5 + heuristic curator
pip install memoirs
# everything except the local LLM (lighter — ~150 MB)
pip install 'memoirs[all]'
# everything + Qwen 3 4B GGUF curator + bge-reranker (~2 GB on first run)
pip install 'memoirs[full]'
memoirs setup # one-shot: download GGUF, init DB, wire MCP into your IDEs
```
`memoirs setup` is idempotent — it auto-detects which curator backends you already have, downloads the missing ones via `huggingface-hub`, then writes MCP config snippets for Claude Code / Cursor / Continue / Cline.
### 2. From source
```bash
git clone https://github.com/misaelzapata/memoirs && cd memoirs
python3 -m venv .venv && source .venv/bin/activate
pip install -e '.[full,dev]' # editable + tests + curator + reranker
memoirs setup
```
### 3. Docker compose (zero-Python install)
```bash
git clone https://github.com/misaelzapata/memoirs && cd memoirs
docker compose up -d
# → API on :8283, MCP via docker exec, persistent volume in ./data
```
### Auto-start on login (Linux, optional)
```bash
bash scripts/install_systemd_user.sh # creates user-level systemd unit
# memoirs now auto-starts on login. status: systemctl --user status memoirs-api
```
### Verify
```bash
memoirs status
# → 12 migrations applied · curator=qwen3 · embedding=sentence-transformers/all-MiniLM-L6-v2
memoirs ingest ~/.claude/projects # pull past chats
memoirs ask "what did we decide about auth?"
```
That's it — the MCP server is wired into your IDEs.
### As an MCP server
```jsonc
// .vscode/mcp.json (or ~/.codex/mcp.json, claude_desktop_config.json, ...)
{
"servers": {
"memoirs": {
"type": "stdio",
"command": "/path/to/.venv/bin/python3",
"args": ["-m", "memoirs", "mcp"]
}
}
}
```
### As a Python library
```python
from memoirs.db import MemoirsDB
from memoirs.engine.memory_engine import assemble_context
db = MemoirsDB("memoirs.sqlite")
ctx = assemble_context(db, "API rate-limit policy", top_k=10)
print(ctx["context"]) # list of compact lines
print(ctx["token_estimate"]) # ~600
```
### As an HTTP API + web dashboard
```bash
memoirs serve --port 8283
# REST: http://localhost:8283/docs
# UI: http://localhost:8283
```
The web UI surfaces every feature live: memory search, timeline, entity graph, conflicts, point-in-time snapshots with side-by-side diff, procedural-memory inspector, and provenance/attribution audit.
---
## Performance & quality
### Internal head-to-head bench (synthetic, 6 engines, 20 queries)
[`scripts/bench_vs_others.py`](scripts/bench_vs_others.py) · artifact: [`bench_results/bench_others_v11_prf.json`](bench_results/bench_others_v11_prf.json)
| engine | MRR | Hit@1 | Hit@5 | R@10 | p50 ms | p95 ms | RAM MB |
|---|---|---|---|---|---|---|---|
| **memoirs** | **1.00** | **1.00** | **1.00** | **1.00** | **21** | **33** | **231** |
| cognee | 0.97 | 0.95 | 1.00 | 1.00 | 974 | 1316 | 1643 |
| mem0 | 0.93 | 0.90 | 1.00 | 1.23\* | 533 | 1465 | 865 |
| memori | 0.93 | 0.85 | 1.00 | 1.00 | 339 | 462 | 1841 |
| langmem | 0.90 | 0.80 | 1.00 | 1.00 | 350 | 616 | 1846 |
| llamaindex | 0.90 | 0.80 | 1.00 | 1.00 | 351 | 627 | 1855 |
\* mem0's R@10 > 1.00 is due to duplicate gold IDs in their response shape — counted as published.
memoirs leads on every retrieval metric and is **16-46× faster** than cloud rivals (no network round trip, no LLM in the hot path). Temporal queries — where memoirs uses `valid_from`/`valid_to` bi-temporal — sweep at MRR 1.00 vs 0.50-0.88 for everyone else.
### Real-world bench: LoCoMo (Snap Research, 1982 QA pairs across 10 long conversations)
[`scripts/eval_locomo.py`](scripts/eval_locomo.py) · artifact: [`bench_results/locomo_full_v1.json`](bench_results/locomo_full_v1.json) · runtime: 873 s
| category | n | MRR | H@1 | H@5 | R@10 | p50 ms |
|---|---|---|---|---|---|---|
| **multi-hop** | 321 | **0.335** | 0.209 | 0.517 | 0.651 | 202 |
| **temporal** | 841 | **0.333** | 0.200 | 0.517 | 0.702 | 182 |
| adversarial | 446 | 0.213 | 0.085 | 0.341 | 0.659 | 140 |
| single-hop | 282 | 0.217 | 0.103 | 0.383 | 0.288 | 209 |
| open-domain | 92 | 0.163 | 0.109 | 0.207 | 0.269 | 223 |
| **total** | **1982** | **0.282** | **0.157** | **0.444** | **0.605** | 182 |
**With bge-reranker** (`MEMOIRS_RERANKER_BACKEND=bge`, `MEMOIRS_PRF=off`) on a 3-conv subset (495 queries, [`locomo_3_rk.json`](bench_results/locomo_3_rk.json)):
| category | baseline MRR | + reranker | uplift |
|---|---|---|---|
| single-hop | 0.182 | **0.494** | **+171%** |
| multi-hop | 0.312 | **0.760** | +144% |
| temporal | 0.298 | **0.558** | +87% |
| adversarial | 0.224 | **0.422** | +89% |
| **TOTAL** | **0.260** | **0.541** | **+108%** |
Cost: ~5 s/query (bge-reranker is a CPU cross-encoder). Trade-off: reranker off = 65× faster, reranker on = 2× quality.
### LongMemEval (xiaowu0162, oracle split, ICLR 2025)
[`scripts/bench_vs_others.py --longmemeval`](scripts/bench_vs_others.py)
| variant | n | MRR | Hit@1 | Hit@5 | R@10 | p50 ms |
|---|---|---|---|---|---|---|
| baseline (PRF on) | 100 | 0.32 | 0.19 | 0.49 | 0.46 | 167 |
| **+ bge-reranker** | 100 | **0.38** | **0.25** | **0.56** | 0.49 | 23,000 |
### Engine internals (warm cache, single thread, 4,061-memory production DB)
| primitive | p50 | p95 | notes |
|---|---|---|---|
| `bm25` (FTS5) | 2.7 ms | 63 ms | lexical only |
| `dense` (sqlite-vec) | 12.8 ms | 4.2 s\* | dense only |
| `hybrid` (BM25 + dense + RRF) | **3.9 ms** | 10 s\* | default |
| `hybrid` + PRF (multi-hop bridge) | 6.5 ms | 12 s\* | `MEMOIRS_PRF=on` |
| `graph` (entity PPR) | 2.1 ms | 333 ms | rarely needed |
| `embed_text` cached | 0.005 ms | — | LRU, ~2,000× speedup |
| migration cold start (1→12) | <200 ms | — | 2 GB DB |
| process cold start (+ Qwen + ST) | 5.7 s | — | one-time |
\* p95 outliers come from the cold-embed path; default hot path stays sub-30 ms.
### Curator quality (LLM JSON adherence, 20-prompt benchmark)
The local curator emits **valid JSON on every contradiction-resolution prompt** with Qwen3-4B. Gemma 2B was the previous default and failed all 20 — that's why Qwen is now the default backend.
### Reproducing the benchmarks yourself
```bash
# 1. Internal head-to-head (fastest — 20 synthetic queries × 6 engines, ~3 min)
MEMOIRS_PRF=on python scripts/bench_vs_others.py \
--engines memoirs,mem0,cognee,memori,langmem,llamaindex \
--top-k 10 \
--out bench_results/bench_others_v11.json \
--md-out bench_results/bench_others_v11.md
# 2. LongMemEval (download the oracle split first, ~15 MB)
mkdir -p ~/datasets/longmemeval
curl -L https://github.com/xiaowu0162/LongMemEval/releases/download/v1.0/longmemeval_oracle.json \
-o ~/datasets/longmemeval/longmemeval_oracle.json
MEMOIRS_PRF=on python scripts/bench_vs_others.py \
--engines memoirs --top-k 10 --longmemeval --longmemeval-limit 100 \
--out bench_results/longmemeval_100.json --md-out bench_results/longmemeval_100.md
# 3. LoCoMo (10 long conversations, 1986 QA pairs)
mkdir -p ~/datasets
curl -L https://raw.githubusercontent.com/snap-research/locomo/main/data/locomo10.json \
-o ~/datasets/locomo10.json
MEMOIRS_PRF=on python scripts/eval_locomo.py \
--locomo ~/datasets/locomo10.json \
--top-k 10 --out bench_results/locomo_full.json
# 4. Same LoCoMo, but with bge-reranker on (slower but +108% MRR)
MEMOIRS_RERANKER_BACKEND=bge python scripts/eval_locomo.py \
--locomo ~/datasets/locomo10.json \
--top-k 10 --out bench_results/locomo_reranked.json
```
Catalog of all 24 catalogued benchmarks (with download recipes + license + estimated effort): [`docs/external_benchmarks_catalog.md`](docs/external_benchmarks_catalog.md).
---
## Quality
The local curator (Qwen 2.5 3B Q4) emits **20/20 valid JSON** on the contradiction-resolution prompt — vs Gemma 2B's 0/20 in our bench. Auto-detected at startup; override with `MEMOIRS_CURATOR_BACKEND={qwen,phi,gemma}`.
```
JSON valid p50 tokens
qwen2.5-3b-Q4 20/20 870 ms 23
phi-3.5-mini-Q4 8/20 raw → 20/20 with parser salvage
gemma-2-2b-Q4 0/20 (chat-template + stop-token interaction)
```
Every curator function has a tolerant parser that salvages truncated / fenced / bare-string JSON, so the system degrades cleanly even when the model misbehaves. Backend selection is auto: Qwen → Phi → Gemma (whichever GGUF is found first in `~/.local/share/memoirs/models/`).
---
## Architecture
```
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ IDE / Agent / Script │
└─────────────────┬────────────────────────┬─────────────────────────┬────────────────┘
│ MCP stdio │ HTTP API + Web UI │ CLI
▼ ▼ ▼
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ memoirs runtime (Python) │
│ │
│ ingest ▶ raw ▶ extract ▶ consolidate ▶ retrieve ▶ serve │
│ chats / conv / Qwen 3 / ADD · UPDATE PRF + MMR /context │
│ events msgs spaCy / MERGE · EXPIRE reranker stream │
│ sources noop ARCHIVE (opt-in) │
│ │
│ ┌──── sleep-time cron (idle housekeeping) ──────────────┐ │
│ │ dedup → link_rebuild → prune → contradictions │ │
│ │ scheduled snapshots (point-in-time backup) │ │
│ └────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────┬──────────────────────────────────────────────┘
│
│ SQLite + sqlite-vec + FTS5
▼
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ memories · candidates · links · entities · summaries · snapshots │
│ │
│ Ebbinghaus strength · Zettelkasten edges · RAPTOR hierarchy │
│ bi-temporal validity · point-in-time backup · provenance_json │
│ (actor · process · decision · reason — migration 012) │
└─────────────────────────────────────────────────────────────────────────────────────┘
```
Engine layers (each independently testable): **raw → extract → consolidate → retrieve → maintain**. PRF (`MEMOIRS_PRF`) and bge-reranker (`MEMOIRS_RERANKER_BACKEND`) plug into retrieve as opt-in env knobs.
**Engine layers** (each independently testable):
1. **Raw** — sources, conversations, messages.
2. **Extract** — Qwen-driven candidates with type validation + secret scanning.
3. **Knowledge graph** — entities, relationships, project context.
4. **Embeddings** — sqlite-vec + sentence-transformers (or fastembed).
5. **Engine** — consolidation, scoring (`importance × confidence × Ebbinghaus(t,S) × usage × signal`), lifecycle, dedup, time-travel, hybrid + graph + RAPTOR retrieval.
6. **MCP / API / UI** — 22 MCP tools, FastAPI with SSE, web inspector at `/ui`.
---
## Features
### Retrieval
- **Hybrid BM25 + dense** with RRF fusion
- **HippoRAG-style multi-hop** via Personalized PageRank over entity + memory graph
- **RAPTOR hierarchical summaries** for long-context queries
- **Streaming SSE** — context streams as it ranks (TTFT < 50 ms)
- **Time-travel** — `as_of=t` returns the corpus state at any past timestamp
- **HyDE / cross-encoder reranker / MMR** — opt-in pipeline stages
### Memory lifecycle
- **Ebbinghaus forgetting curve** — `R = e^(-Δt / (S·24h))`, strength × 1.5 per access (cap 64)
- **Zettelkasten linking** — bidirectional `memory_links` with 4 modes (absolute / topk / adaptive / zscore)
- **EXPIRE/ARCHIVE generation** — heuristic + LLM curator decide when to retire memories
- **Sleep-time consolidation** — cron job runs dedup/prune/contradictions on idle
- **Versioning** — bi-temporal `valid_from` / `valid_to` per memory
### Privacy & ops
- **PII redaction** with Microsoft Presidio (optional) + 11 always-on secret detectors
- **Per-memory ACL** with `private / shared / org / public` visibility tiers
- **Encryption-at-rest** via SQLCipher (`MEMOIRS_ENCRYPT_KEY`)
- **GDPR export/import** as portable zip bundles with sha256 manifests
- **Structured JSON logging** + OpenTelemetry traces (opt-in)
- **Versioned migrations** with `up()` / `down()` and rollback
### Integrations
- **MCP** stdio server (Claude Desktop, Claude Code, Cursor, Continue, Cline, Codex)
- **HTTP API** with FastAPI + SSE streaming
- **Inspector UI** at `/ui` — timeline, graph, search, conflict-resolution, edit/pin/forget
- **CLI** with 30+ subcommands
- **Eval harness** — LongMemEval / LoCoMo / synthetic suites with precision@k / recall@k / MRR
---
## CLI tour
```bash
# Ingestion
memoirs ingest <path|url> # any supported format, auto-detected
memoirs ingest --kind claude-export memories.zip
memoirs watch ~/.claude/projects # real-time
memoirs daemon start # background watcher + extractor + sleep cron
# Retrieval & inspection
memoirs ask "..." # one-shot context query
memoirs why <memory_id> --query "..." # provenance trace
memoirs trace <conversation_id> # source → messages → candidates → memories
memoirs graph entities --html # interactive entity visualization
memoirs ui --port 8284 # web inspector
# Lifecycle
memoirs maintenance # recompute scores + expire
memoirs cleanup # dedup near-duplicates
memoirs links rebuild --mode topk # rebuild Zettelkasten edges
memoirs raptor build # construct hierarchical summaries
memoirs sleep run-once # one cron tick
# Privacy & data
memoirs export --user-id alice --redact-pii --out alice.zip
memoirs import alice.zip --mode merge
memoirs db encrypt --key 'passphrase'
# Eval
memoirs eval --suite synthetic_basic --modes hybrid,bm25,dense
```
Run `memoirs --help` for the full surface (35 subcommands).
---
## MCP tools (22)
```
Raw ingest_event • ingest_conversation • status
Extraction extract_pending • consolidate_pending • audit_corpus
record_tool_call • consolidate_with_gemma
Engine run_maintenance • get_context • summarize_thread
search_memory • add_memory • update_memory
score_feedback • explain_context • forget_memory
list_memories
Graph index_entities • get_project_context • list_projects
Ops event_stats • export_user_data • import_user_data
```
`mcp_get_context` is the workhorse — every other tool exists to make its output better. Returns ~600–1,500 tokens of ranked, conflict-resolved memory with optional `provenance_chain` for explainability.
---
## Configuration
Everything is env-var driven so the same binary works in dev, daemon, and CI.
```bash
# Storage
MEMOIRS_DB=path/to.sqlite # default .memoirs/memoirs.sqlite
MEMOIRS_ENCRYPT_KEY=<passphrase> # enable SQLCipher
MEMOIRS_SQLITE_MMAP_MB=256 # default 256
MEMOIRS_SQLITE_CACHE_MB=64 # default 64
# Curator (LLM) — Qwen 2.5 3B is the default when its GGUF is present.
MEMOIRS_CURATOR_BACKEND=qwen|phi|gemma # default: auto-detect (qwen > phi > gemma)
MEMOIRS_CURATOR_MODEL=/path/to.gguf
MEMOIRS_GEMMA_THREADS=20 # legacy name, applies to whichever curator is loaded
MEMOIRS_GEMMA_GPU_LAYERS=99 # Vulkan offload
# Retrieval
MEMOIRS_RETRIEVAL_MODE=hybrid_graph # dense|bm25|hybrid|graph|hybrid_graph|raptor|hybrid_raptor
MEMOIRS_RETRIEVAL_GEMMA=off # off | on (slow but smarter conflict resolution)
MEMOIRS_RETRIEVAL_GEMMA_MAX=2 # cap LLM calls per query
MEMOIRS_HYDE=off # query expansion
MEMOIRS_RERANKER_BACKEND=none|bge # cross-encoder rerank top-N
MEMOIRS_MMR=on
MEMOIRS_MMR_LAMBDA=0.7
# Embedding
MEMOIRS_EMBED_BACKEND=sentence_transformers|fastembed
# Privacy
MEMOIRS_REDACT=on|off|strict # PII + secret scanning at ingest
MEMOIRS_USER_ID=alice # multi-tenant scope
MEMOIRS_NAMESPACE=work
# Observability
MEMOIRS_LOG_FORMAT=json|text
MEMOIRS_LOG_LEVEL=INFO
MEMOIRS_OTEL_ENDPOINT=http://otelcollector:4317
# Lifecycle
MEMOIRS_ZETTELKASTEN=on|off
MEMOIRS_GEMMA_CURATOR=on|off|auto # consolidation curator
```
Full list in [`docs/configuration.md`](docs/configuration.md).
---
## Comparison
| | Local-first | Hybrid retrieval | Multi-hop bridge | Temporal queries | Native MCP | Auto-curate | Eval harness |
|---|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
| **memoirs** | ✅ SQLite | ✅ BM25+dense+RRF | ✅ PRF + PPR | ✅ bi-temporal `as_of=t` | ✅ 29 tools | ✅ Qwen 3 / Phi / Gemma | ✅ LongMemEval, LoCoMo, synthetic |
| **Mem0** | ⚠️ SaaS-default | ✅ | ❌ | partial | ⚠️ wrapper | ✅ OpenAI/Anthropic | ⚠️ in-house only |
| **Cognee** | ⚠️ Postgres | ✅ | ✅ ontology graph | ❌ | ❌ | ✅ LiteLLM | ⚠️ in-house |
| **Letta** (MemGPT) | ⚠️ Postgres | ✅ | ❌ | ❌ | ⚠️ wrapper | ✅ OpenAI | partial |
| **Zep / Graphiti** | ❌ SaaS | ✅ | ✅ entity graph | ✅ | ⚠️ wrapper | ✅ OpenAI | ✅ DMR + LongMemEval |
| **LangMem** | ✅ in-process | ⚠️ ad hoc | ❌ | ❌ | ⚠️ wrapper | ✅ procedural | ❌ |
| **LlamaIndex** | ✅ in-process | ✅ | ⚠️ via plugins | ❌ | ⚠️ wrapper | ❌ DIY | ❌ |
| **Memori** | ✅ SQLite | ✅ | ❌ | ❌ | ❌ | ✅ heuristic | ❌ |
---
## Project status
- 693 tests passing • coverage 60% • 9 schema migrations • 35 CLI subcommands • 22 MCP tools
- Tested on a real 4,196-memory corpus: p50 6.2 ms hybrid_graph
- LongMemEval / LoCoMo head-to-head vs other engines is the next milestone
---
## Documentation
- [`docs/configuration.md`](docs/configuration.md) — every env var explained
- [`docs/architecture.md`](docs/architecture.md) — schema, layers, data flow
- [`docs/benchmarks.md`](docs/benchmarks.md) — methodology + raw numbers
- [`docs/mcp.md`](docs/mcp.md) — MCP tool reference
- [`docs/cli.md`](docs/cli.md) — CLI reference
---
## Contributing
Issues and PRs welcome. Run the suite first:
```bash
.venv/bin/pytest tests/ -q # 693 tests, ~90 s
bash scripts/coverage.sh # coverage report → .coverage_html/
```
`CONTRIBUTING.md` (TODO) covers the migration cookbook, test conventions, and how to add a new MCP tool.
---
## License
MIT — see [`LICENSE`](LICENSE).
---
## Screenshots
Built against the seeded demo DB (`scripts/seed_demo_db.py`) — 51 memorias spanning every memory type, plus a snapshot baseline so the diff is non-empty.
| | |
|---|---|
| **Dashboard** — corpus stats + doughnut of memory types + procedural policies |  |
| **Memories** — full list with type pills + provenance |  |
| **Procedural** — `?type=procedural` filter |  |
| **Search** — BM25 + dense + RRF retrieval |  |
| **Timeline** — bi-temporal view, filter by type |  |
| **Conflicts** — semantic dups across types |  |
| **Snapshots** — list + create + auto-cadence stat |  |
| **Snapshot diff** — side-by-side, added/removed/changed in colored cards |  |
To regenerate:
```bash
python scripts/seed_demo_db.py --out data/demo/memoirs_demo.sqlite
memoirs --db data/demo/memoirs_demo.sqlite serve --port 8283 &
bash scripts/take_screenshots.sh
```
---
MCP 配置
以下是此 MCP Server 的配置信息，可直接复制到 Cursor 或其他 MCP 客户端中使用。
mcp.json
连接信息
猜你喜欢
暂无推荐服务器