longmem
Cross-project memory for AI coding assistants.
Stop solving the same problems twice.
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 projectWhy 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 |
Related MCP server: MCP Memory Server
Quickstart
1. Install
pipx install longmem2. Setup — checks Ollama, pulls the embedding model, writes your IDE config
longmem init3. Activate in each project — copies the rules file that tells the AI how to use memory
cd your-project
longmem install4. 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.jsonOption 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 |
| One-time setup: Ollama check, model pull, writes IDE config |
| Copy rules into the current project |
| Config, Ollama reachability, entry count, DB size |
| Dump all entries to JSON — backup or share |
| Load a JSON export — onboard teammates or migrate machines |
| 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_KEYInstall 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_KEYNo 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 |
|
3. Project detail | One repo only |
|
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 akeyword_matchflag 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 |
| GitHub Actions, Jenkins, GitLab CI, build failures |
| Docker, Kubernetes, Helm, OOM kills |
| Terraform, Pulumi, CDK, IaC drift |
| AWS/GCP/Azure SDK, IAM, quota errors |
| DNS, TLS, load balancers, timeouts, proxies |
| Logging, metrics, tracing, Prometheus, Grafana |
| OAuth, JWT, RBAC, secrets, CVEs |
| Airflow, Prefect, Dagster, ETL, data quality |
| GPU/CUDA, distributed training, OOM |
| vLLM, Triton, inference latency, batching |
| MLflow, W&B, DVC, reproducibility |
| Chunking, embedding, retrieval, reranking |
| Rate limits, token cost, prompt engineering |
| Pinecone, Weaviate, Qdrant, LanceDB |
| LangChain, LlamaIndex, tool-calling, agent memory |
| SQL/NoSQL, migrations, slow queries |
| REST, GraphQL, gRPC, versioning |
| Race conditions, event loops, deadlocks |
| Version conflicts, packaging, lock files |
| Profiling, memory leaks, caching |
| Flaky tests, mocks, integration vs unit |
| Design patterns, service boundaries, refactoring |
| 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 pytestGood 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
Available Tools
11 toolsadd_edge_caseA
Record a context where a cached solution didn't work as-is.
Call this when search_similar returned a match but it needed modification to work in the current project. The edge case is appended to the entry so future suggestions include the caveat.
Inputs: entry_id (required), edge_case (required). Returns: result
confirm_solutionA
Auto-save a confirmed solution using context from the last search_similar call.
Call this after solving a problem instead of save_solution — you only need to provide the solution text. Problem description, category, tags, and language are taken automatically from the last search_similar call.
If save_solution was already called manually this session, this is a no-op (no duplicate will be created).
Inputs: solution (required), project. Returns: result
correct_solutionA
Fix a specific piece of text in an already-saved solution.
Call this when the user corrects a name, term, or detail that was saved
incorrectly — for example 'it's not called Paperless-NGX, it's Papertagging'.
Replaces all occurrences of find with replace in the solution text.
Use enrich_solution to add new context. Use correct_solution to fix wrong text.
Inputs: entry_id (required), find (required), replace (required). Returns: result
delete_solutionA
Permanently delete a saved entry.
Use this to remove entries that were saved incorrectly, contain wrong information that can't be fixed with correct_solution, or are no longer relevant. This cannot be undone.
Inputs: entry_id (required). Returns: result
enrich_solutionA
Append new context to an already-saved solution.
Call this when a conversation reveals additional details AFTER a solution was already saved — for example, a follow-up clarification that makes the solution more reusable across projects.
This is NOT for failures (use add_edge_case for those). This is for enrichment: new facts, patterns, or context that improve the answer.
Inputs: entry_id (required), context (required). Returns: result
list_recentA
List the most recently saved memory entries.
Use this to audit what has been saved — for example, to find a recently saved entry whose id is not in context. Results are ordered newest-first.
Inputs: limit. Returns: result
rebuild_indexA
Rebuild the vector search index for faster similarity search.
LanceDB falls back to brute-force scan when the table has fewer than 256 rows. Once you have 256+ entries, call this once to build an ANN index — subsequent searches will be significantly faster.
Safe to call at any time; existing data is not modified.
Inputs: no named parameters. Returns: result
save_solutionA
Save a problem/solution pair to the cross-project memory.
Call this after successfully solving a problem so future sessions — in any project — can find and reuse the solution. Returns the entry ID which can be passed to add_edge_case later.
Inputs: problem (required), solution (required), category (required), project, +2 more. Returns: result
search_by_projectA
List saved entries for a specific project.
Use this at the start of a new conversation when you need to find a project-specific entry to correct or enrich but no entry_id is in context. Returns entry ids, problems, and solutions so you can pick the right one and pass its id to correct_solution or enrich_solution.
Inputs: project (required), query, limit. Returns: result
search_similarA
Search the cross-project memory for solutions similar to the current problem.
Call this FIRST before reasoning about a problem from scratch. If similarity ≥ threshold a cached solution is returned — check edge_cases to see if any known limitations apply to the current context. If no match is found, solve normally and then call confirm_solution.
Inputs: problem (required), category, tags, language, +1 more. Returns: result
statsA
Return database statistics: total entries, breakdown by category, and date range.
Useful for understanding the size and composition of the memory store, and for deciding when to call rebuild_index (threshold: 256+ entries).
Inputs: no named parameters. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
Each tool has a clearly distinct purpose: edge cases, confirmation, correction, deletion, enrichment, listing, indexing, saving, project search, similarity search, and statistics. There is no ambiguity between them.
All tools use snake_case with a verb_noun pattern (e.g., add_edge_case, confirm_solution). The exception 'stats' is a common shorthand and still fits the pattern as a noun. Overall highly consistent.
With 11 tools, the set is well-scoped for a memory management system. It covers all necessary operations without being bloated or insufficient.
The surface covers saving, retrieving, updating, deleting, and searching solutions, plus edge cases and enrichment. Minor gaps exist, such as no direct 'get solution by ID' tool, but this can be worked around via search methods.
Maintenance
Related MCP Connectors
Your own memory for every AI: Claude, ChatGPT, Cursor, Zed and more. Brings in 20+ sources.
Project memory for every AI. Pick up where you left off in Claude, ChatGPT or Cursor, with a team.
Persistent memory for Claude Code and Cursor. Stop re-explaining your project every session.
One memory for your AI agents and your team: Claude, ChatGPT, Cursor and your agents share it.
Related MCP Servers
- AlicenseAqualityCmaintenanceProvides persistent memory and a codebase knowledge graph for AI coding assistants, enabling shared context across multiple tools like Claude, Cursor, and ChatGPT, with significant token reduction.526 npmMIT
- AlicenseNot gradedqualityDmaintenanceProvides long-term memory for AI coding agents, enabling them to remember, search, and organize information across sessions and platforms like Claude Code, ChatGPT, and Cursor.8MIT
- AlicenseCqualityBmaintenanceGives AI coding assistants persistent project memory and semantic code search, running fully locally with no API keys required.56MIT
- AlicenseAqualityDmaintenancePersistent memory + intelligent context for Claude Code. Save and reuse solutions, decisions, and patterns across projects.437MIT