MemoryHub

A centralized memory service for AI agents.

Personal AI agents (Claude Code, etc.) each build up knowledge as local Markdown files, in isolation. MemoryHub pools those per-user memories into one shared, searchable knowledge base: it ingests raw Markdown from each agent, embeds it for semantic search, and runs a synthesizer that produces per-user and cross-user summaries.

What it does

• Collects memories from many users and agents over a small JSON API.
• Searchable instantly — every memory is indexed the moment it's written, so it's immediately available. No batch or reprocess step.
• Finds the right memory — search handles both meaning-based and exact keyword queries, so results hold up whether you describe an idea or quote a term.
• Synthesizes — generates per-user and cross-user summaries from the pooled memories as they come in.
• Human-readable storage — memories stay as plain Markdown files you can read, edit, and version directly.
• Bring your own models — chat and embeddings are configured separately, so you can mix providers (and swap in a stub for tests).

It's a single self-contained service.

Quick start

# Run the server (defaults: 0.0.0.0:8000, data under ~/.memoryhub)
cargo run -p memoryhub -- --host 127.0.0.1 --port 8000

# Write a memory
curl -X POST http://127.0.0.1:8000/v1/memories/write \
  -H 'Content-Type: application/json' \
  -d '{"username":"alice","agent_id":"a1","filename":"role.md","content":"# Role\nBackend engineer"}'

# Search
curl -X POST http://127.0.0.1:8000/v1/search \
  -H 'Content-Type: application/json' \
  -d '{"username":"alice","agent_id":"a1","query":"who works on the backend?"}'

Configuration is a config.toml under the data directory (chat/embedding provider, models, chunking, paths). The data directory is chosen by, in order: --base-dir > $MEMORYHUB_HOME > ~/.memoryhub.

License

This project is licensed under MIT.