MCP 服务器

longmem

Hybrid semantic + keyword memory across all your projects. Works with Cursor, Claude Code, and your team.

README

Cross-project memory for AI coding assistants.
Stop solving the same problems twice.

</div>

Your AI solves the same bug in a different project six months later. Writes the same boilerplate. Explains the same pattern. You already knew the answer.

longmem gives your AI a persistent memory that works across every project and every session. Before reasoning from scratch, it searches what you've already solved. After something works, it saves it. The longer you use it, the less you repeat yourself.

You describe a problem
        │
        ▼
  search_similar()
  ┌─────────────────────────────────────────────────────┐
  │  1. pre-filter by category (ci_cd / auth / db / …)  │
  │  2. semantic search  (Ollama or OpenAI embeddings)   │
  │  3. keyword search   (SQLite FTS5 exact match)       │
  │  4. merge + rank results                             │
  └─────────────────────────────────────────────────────┘
        │                          │
   score ≥ 85%               score < 85%
        │                          │
        ▼                          ▼
  cached solution           AI reasons from scratch
  + edge cases                      │
  + team knowledge               "it works"
  (any project)                     │
                                    ▼
                          confirm_solution()
                          saved once — surfaces
                          from every future project

Why longmem

	longmem	others
Cost	Free — local Ollama embeddings	Requires API calls per session
Privacy	Nothing leaves your machine	Sends observations to external APIs
Process	Starts on demand, no daemon	Background worker + open port required
IDE support	Cursor + Claude Code	Primarily one IDE
Search	Hybrid: semantic + keyword (FTS5)	Vector-only or keyword-only
Teams	Export / import / shared DB path / S3	Single-user
License	MIT	AGPL / proprietary

Quickstart

1. Install

pipx install longmem

2. Setup — checks Ollama, pulls the embedding model, writes your IDE config

longmem init

3. Activate in each project — copies the rules file that tells the AI how to use memory

cd your-project
longmem install

4. Restart your IDE. Memory tools are now active on every chat.

Need Ollama? Install from ollama.com, then ollama pull nomic-embed-text. Or use OpenAI — see Configuration.

How it works

longmem is an MCP server. Your IDE starts it on demand. Two rules drive the workflow:

Rule 1 — search first. Before the AI reasons about any bug or question, it calls search_similar. If a match is found (cosine similarity ≥ 85%), the cached solution is returned with any edge-case notes. Below the threshold, the AI solves normally.

Rule 2 — save on success. When you confirm something works, the AI calls confirm_solution. One parameter — just the solution text. Problem metadata is auto-filled from the earlier search.

The rules file (longmem.mdc for Cursor, CLAUDE.md for Claude Code) wires this up automatically. No manual prompting.

AI forgot to save? Run longmem review — an interactive CLI to save any solution in 30 seconds.

Cold start — getting value from day one

longmem is most useful once it has entries. The fastest way to seed it:

Option 1 — review as you go. After every solved problem this week, run longmem review and describe what you fixed. Ten entries is enough to feel the difference.

Option 2 — team import. If a teammate already has entries, they export and you import:

# teammate
longmem export team_knowledge.json

# you
longmem import team_knowledge.json

Option 3 — shared DB. Set db_path (or db_uri for S3/cloud) to the same location for the whole team. Every save is instantly available to everyone.

CLI

Command	What it does
`longmem init`	One-time setup: Ollama check, model pull, writes IDE config
`longmem install`	Copy rules into the current project
`longmem status`	Config, Ollama reachability, entry count, DB size
`longmem export [file]`	Dump all entries to JSON — backup or share
`longmem import <file>`	Load a JSON export — onboard teammates or migrate machines
`longmem review`	Manually save a solution when the AI forgot

longmem with no arguments starts the MCP server (used by your IDE).

Configuration

Config lives at ~/.longmem/config.toml. All fields are optional — defaults work with a local Ollama instance.

Switch to OpenAI embeddings

embedder       = "openai"
openai_model   = "text-embedding-3-small"
openai_api_key = "sk-..."   # or set OPENAI_API_KEY

Install the extra: pip install 'longmem[openai]'

Team shared database

Point every team member's config at the same path:

# NFS / shared drive
db_path = "/mnt/shared/longmem/db"

Or use cloud storage:

# S3 (uses AWS env vars)
db_uri = "s3://my-bucket/longmem"

# LanceDB Cloud
db_uri = "db://my-org/my-db"
lancedb_api_key = "ldb_..."   # or set LANCEDB_API_KEY

No shared mount? Use longmem export / longmem import to distribute a snapshot.

Team knowledge base

Save facts that are true across your whole stack under project="shared" so they surface from any repo:

save_solution(
  problem="why oauth2-proxy uses port 4181 not default 4180",
  solution="General: 4180 is the oauth2-proxy default. 4181 means something else already occupies 4180.\n\nThis team's setup: Sinfonia always runs on 4180. Every other project uses 4181+ by convention.",
  project="shared",
  category="networking",
  tags=["oauth2-proxy", "ports", "nginx"]
)

search_similar searches all projects — a shared entry surfaces automatically from any repo without needing search_by_project.

Three-layer solution format — write solutions so they work for anyone who finds them:

Layer	Scope	How to save
1. General pattern	Universal — any team	always include in solution text
2. Team-wide fact	Your whole stack	`project="shared"`
3. Project detail	One repo only	`project="<repo>"` + `enrich_solution`

Tuning

similarity_threshold = 0.85   # minimum score to surface a cached result (default 0.85)
duplicate_threshold  = 0.95   # minimum score to block a save as a near-duplicate (default 0.95)

MCP tools

The server exposes 11 tools. The two you interact with most:

search_similar — semantic + keyword hybrid search. Returns ranked matches with similarity scores, edge cases, and a keyword_match flag when the hit came from exact text rather than vector similarity.
confirm_solution — saves a solution with one parameter. Problem metadata auto-filled from the preceding search.

Full list: save_solution, correct_solution, enrich_solution, add_edge_case, search_by_project, delete_solution, rebuild_index, list_recent, stats.

Call rebuild_index once you reach 256+ entries to compact the database and build the ANN index for faster search.

Category reference

Categories pre-filter before vector search — keeps retrieval fast at any scale.

Category	Use for
`ci_cd`	GitHub Actions, Jenkins, GitLab CI, build failures
`containers`	Docker, Kubernetes, Helm, OOM kills
`infrastructure`	Terraform, Pulumi, CDK, IaC drift
`cloud`	AWS/GCP/Azure SDK, IAM, quota errors
`networking`	DNS, TLS, load balancers, timeouts, proxies
`observability`	Logging, metrics, tracing, Prometheus, Grafana
`auth_security`	OAuth, JWT, RBAC, secrets, CVEs
`data_pipeline`	Airflow, Prefect, Dagster, ETL, data quality
`ml_training`	GPU/CUDA, distributed training, OOM
`model_serving`	vLLM, Triton, inference latency, batching
`experiment_tracking`	MLflow, W&B, DVC, reproducibility
`llm_rag`	Chunking, embedding, retrieval, reranking
`llm_api`	Rate limits, token cost, prompt engineering
`vector_db`	Pinecone, Weaviate, Qdrant, LanceDB
`agents`	LangChain, LlamaIndex, tool-calling, agent memory
`database`	SQL/NoSQL, migrations, slow queries
`api`	REST, GraphQL, gRPC, versioning
`async_concurrency`	Race conditions, event loops, deadlocks
`dependencies`	Version conflicts, packaging, lock files
`performance`	Profiling, memory leaks, caching
`testing`	Flaky tests, mocks, integration vs unit
`architecture`	Design patterns, service boundaries, refactoring
`other`	When nothing above fits

Contributing

Contributions are very welcome — this project grows with the community that uses it.

Whether it's a bug fix, a new feature, better docs, or just sharing your use case — all of it helps. If you're unsure whether an idea fits, open an issue first and we'll figure it out together.

Getting started:

git clone https://github.com/marerem/longmem
cd longmem
uv sync --group dev
uv run pytest

Good first contributions:

New category suggestions
Edge cases you hit in real projects
IDE integrations (JetBrains, VS Code, Neovim, etc.)
Better error messages
Seed datasets — export your own entries and share them as a starter pack

Ways to contribute without code:

Star the repo if you find it useful
Share it with your team
Open an issue if something is confusing — unclear UX is a bug

License

MIT — see LICENSE.

mcp-name: io.github.marerem/longmem