Stop rebuilding RAG pipelines. Extremis gives any AI agent layered, learning memory in two lines of config — local SQLite, Postgres, Chroma, or Pinecone.
pip install extremis
Every conversation logs durably. Every recalled memory gets scored. Every night the logs distil into structured knowledge.
Append to a fsync'd JSONL log and write to the episodic store in one call. Durable before any embedding happens.
Semantic search ranked by cosine × RL score × recency. Identity and procedural memories always included. Every result ships with a plain-English reason.
Mark recalled memories useful (+1) or harmful (-1). Negative signals apply 1.5× weight — same asymmetry human threat-memory uses.
The "dream pass" — reads new log entries, calls Claude Haiku to extract durable semantic and procedural facts, advances the checkpoint. Safe to re-run.
Requires Python 3.11+. No API key needed for the core library.
# Core — local SQLite, sentence-transformers (no API key) pip install extremis # + MCP server for Claude Desktop / Code pip install "extremis[mcp]" # + Postgres, Chroma, Pinecone, OpenAI embeddings, hosted server pip install "extremis[all]"
from extremis import Extremis, MemoryLayer mem = Extremis() # stores in ~/.extremis/ by default # Store mem.remember("User is building a WhatsApp AI", conversation_id="c1") mem.remember_now("Always ask about deadlines first", layer=MemoryLayer.PROCEDURAL) # Recall — every result explains why it ranked here results = mem.recall("what is the user building?") for r in results: print(r.memory.content) print(r.reason) # "similarity 0.87 · score +2.0 · used 5× · 3d old" # Reinforcement — good memories surface first over time mem.report_outcome([r.memory.id for r in results[:2]], success=True) # Try the demo # $ extremis-demo
Most memory systems are cosine search over a flat vector DB. Extremis is different in five ways.
Recency decay + 1.5× asymmetric RL weighting. Unimportant things fade. Mistakes stick harder. The same pattern human threat-memory uses.
r.reason — "similarity 0.91 · score +4.0 · used 8× · 3d old". Auditable for enterprises. Debuggable for everyone.
Multiple agents share one namespace. Research agent stores, writing agent recalls. Agent teams ship as one brain.
SQLite locally. Postgres in staging. Pinecone in production. One env var. Migrate between any two with extremis-migrate.
Entities, relationships, attributes. Answers "who does Alice work for" — the structural question vectors can't.
Plug into Claude Desktop or Claude Code with two lines of config. Nine tools, no code required.
| Feature | Extremis | Mem0 | LangChain | Raw Pinecone |
|---|---|---|---|---|
| RL-scored retrieval | ✓ | ✗ | ✗ | ✗ |
| Memory explains itself | ✓ | ✗ | ✗ | ✗ |
| Knowledge graph | ✓ | ✗ | ✗ | ✗ |
| Backend-agnostic | ✓ 4 backends | Cloud only | Manual | — |
| MCP server | ✓ 9 tools | ✗ | ✗ | ✗ |
| Self-hostable | ✓ | ✗ | ✓ | ✓ |
| Migration CLI | ✓ | ✗ | ✗ | — |
| Open source (MIT) | ✓ | Partial | ✓ | ✗ |
Claude Desktop and Claude Code supported out of the box.
# 1. Install pip install "extremis[mcp]" # 2. Add to Claude Desktop config # ~/Library/Application Support/Claude/ # claude_desktop_config.json { "mcpServers": { "extremis": { "command": "extremis-mcp" } } } # 3. Restart Claude Desktop
No cloud. No tracking. No data leaving your machine unless you explicitly choose it.
All memories, logs, and knowledge graph data are written to
~/.extremis/
on your own machine. No account required. No sign-up. Nothing sent anywhere.
The conversation log is a JSONL file on your disk. The memory store is a SQLite database. Both are readable with any text editor. You can inspect, export, or delete them at any time.
The optional memory_consolidate tool
sends recent conversation log entries to Claude Haiku (via your own Anthropic API key) to extract
structured memories. You control this — it only runs when you explicitly call it,
and only uses your key, billed to you directly.
Switch to Postgres, Chroma, or Pinecone when you need scale — or keep SQLite forever. If you use an external backend, your data goes there under your own credentials. Extremis itself never sees it.
Extremis collects no usage data, sends no analytics, and has no home-call behaviour. It's a library that runs entirely in your process. You can verify this — it's open source.
Need a REST API for a team? Run extremis-server
on your own infrastructure. The hosted server code is open source and ships with a
Docker Compose file. We offer a managed cloud option for convenience, not necessity.