Hybrid recall, not just vectors
Vector similarity and BM25 fused with reciprocal rank, then reranked. A name or an exact phrase is found by the half of the search that is good at names.
MEMORY FOR AI AGENTS · OVER MCP
CaBrain
A memory organ for AI agents. Long-term recall and an entity graph that any agent reaches over MCP — so what one session learns, the next one already knows.
claude — cabrain
❯ memory_recall{ namespace: "fadymondy", query: "why is /docs scoped?" }hybrid recall · vector + BM25 · RRF · reranked✓ 3 memories, 41ms — hot tier❯ memory_retain{ content: "docs are per-site", source_kind: "claude_code" }salience gate → retained · ADD✓ consolidation scheduled
togo install togo-framework/brain- MCP tools for any client
- Postgres · pgvector · BM25
- Redis L1 cache
- Hot and cold tiers
- Providers ship as plugins
- Live at cabrain-app.fadymondy.com
Features
Hot and cold, like a brain
A hippocampal hot tier answers fast; a cortical cold tier keeps everything. Salience-gated consolidation moves what matters between them while you are not looking.
Reconsolidation on recall
Remembering something changes it — recall strengthens a memory and lets it be corrected, rather than leaving a stale fact to be retrieved for ever.
Any client, one protocol
Claude Desktop and Code, Codex, Gemini and Cursor connect through one CLI that is both the MCP server and its installer.
Why CaBrain?
Four ways an agent can know something, from the shortest-lived to the longest.
conversation history one session lost when the context resets
↓
traditional RAG documents retrieved, but the agent cannot write back
↓
persistent memory what it learned retained now, recalled next session
↓
CaBrain memory + graph hybrid recall · entities · tiers · over MCPConversation history
Everything the agent knows lives in the context window, and goes when the session ends or the window fills.
Traditional RAG
Retrieval over documents someone indexed. Useful for reference material; the agent reads it but does not add to it.
Persistent memory
The agent retains decisions and facts as it works, and recalls them in a later session — so the second session starts where the first ended.
CaBrain
Persistent memory with an entity graph, vector + BM25 recall and hot/cold tiers, behind MCP tools. Connectors still ingest text, markdown, GitHub, SQL and crawled pages — RAG sits inside it, not against it.
What an agent can call
- memory_recall
- memory_retain
- memory_get
- memory_edit
- memory_forget
- graph_neighbors
- graph_traverse
- brain_chat
What it is built on
01
Store
- togo-postgres
- pgvector
- VectorChord
- BM25
02
Cache
- Redis
03
Providers
- brain-tei
- brain-cognee
Get started
- 01
Install the plugin
togo install togo-framework/brain - 02
Connect your AI client
cabrain connect claude-code - 03
Give it a namespace
cabrain namespace create my-project
Status
Live. The console and the MCP API run at cabrain-app.fadymondy.com on togo-postgres with VectorChord and BM25, embedding and reranking through TEI — retain, recall and the entity graph are in daily use.
Community
GitHubmaintained
More from the ecosystem
Give your agents a memory.
Open the console to create a brain, mint a scoped token and connect Claude Code, Cursor or any MCP client.