Semantic memory
HNSW vector search over local embeddings for fast recall from agent context.
Memory for AI agents that outlives the session.
ironmem is an open-source Rust MCP server that gives Codex, Claude Code, and any MCP client a private semantic memory: search, diary, taxonomy, and knowledge-graph tools over one local SQLite store. No cloud, no Python — a single binary that starts serving in under 25 ms.
One local server, one database, and a tool surface agents can share.
HNSW vector search over local embeddings for fast recall from agent context.
ONNX sentence embeddings live inside the workspace through ironrace-embed.
Expose status, search, diary, graph, taxonomy, and collaboration tools over stdio.
Structured facts and semantic drawers can be queried together as memory grows.
Codex and Claude Code read and write the same SQLite store by default.
Bounded Claude↔Codex planning and coding sessions run over the shared store.
The source quickstart builds the binary, runs setup, then starts the trusted MCP server.
git clone https://github.com/ironrace/ironmem.git
cd ironmem
scripts/install-ironmem.sh
~/.ironrace/bin/ironmem setup
IRONMEM_MCP_MODE=trusted ~/.ironrace/bin/ironmem serve
macOS (arm64) and Linux (x86_64) tarballs ship with every tagged release, with SHA-256 checksums.
Download a release →Validate stdio startup without downloading the model:
python3 scripts/mcp_smoke_test.py --binary ~/.ironrace/bin/ironmem
Point Codex at the installed binary and set trusted mode when write tools are needed.
Open Codex setup →ironmem is built for local agent workflows where memory should outlive one chat window.
Use ironmem as a local MCP server for shared search, status, graph, and write tools.
Setup guide →Plugin packaging is included so both harnesses can use the same memory database.
Collab guide →ironrace-core and ironrace-embed provide the shared index and embedding layers.
Browse crates →ironmem is developed in the open under the Apache-2.0 license.