agentcache

A Python REST + WebSocket + MCP cache server for AI agents, backed by SQLite.

agentcache gives coding agents (Claude Code, Cursor, Cline, Kiro, Antigravity, …) a persistent, searchable memory of what happened in a project — observations, sessions, long-term memories, and a folder-scoped knowledge graph — with hybrid BM25 + vector search, WebSocket live updates, and a Model Context Protocol (MCP) stdio bridge.

• Server: Flask + flask-sock (REST + WebSocket) on a single port
• Storage: SQLite in WAL mode (zero external services)
• Search: BM25 + optional local embeddings (sentence-transformers) or Hugging Face inference
• MCP: stdio bridge so any MCP-compatible client can observe / remember / search natively
• CLI: agentcache serve|worker|migrate|export|context|connect

Status: Beta (Development Status :: 4 - Beta). API is stable enough for daily use; internals may still shift before 1.0.

Install

pip install agentcache-core

The package installs as agentcache-core on PyPI but imports and runs as agentcache (import agentcache, agentcache serve). The plain agentcache name on PyPI belongs to an unrelated project.

With optional local embeddings (adds sentence-transformers, ~1 GB of model weights on first use):

pip install "agentcache-core[local-embeddings]"

Requires Python 3.10+.

Quickstart

Start the server:

agentcache serve --port 3111

Record an observation (any HTTP client works):

curl -X POST http://localhost:3111/agentcache/observe \
  -H "Content-Type: application/json" \
  -d '{
        "folderPath": "/abs/path/to/project",
        "agentId": "coder-1",
        "text": "Refactored auth into a decorator; see routes/auth.py",
        "type": "refactor"
      }'

Search across a folder:

curl "http://localhost:3111/agentcache/folder/observations?folderPath=/abs/path/to/project&q=auth"

Health check:

curl http://localhost:3111/agentcache/health

Generate a live context file for your IDE:

cd /abs/path/to/project
agentcache context --watch     # writes .agentcache_context.md and re-writes on change

Programmatic use

from agentcache import create_app, StateKV, remember

app = create_app()                 # Flask app, ready for WSGI or app.run()
kv  = StateKV()                    # direct KV access to the SQLite store
remember(kv, {"title": "auth decorator", "content": "…", "agentId": "coder-1"})

Top-level exports: create_app, StateKV, KV, ObservationStore, ObservationEvents, SearchService, remember, folder_graph_build, health_check, run_connect.

Wire it into your agent

agentcache connect claude-code    # or: cursor, cline, kiro, antigravity, codex, hermes, vscode
agentcache connect --all          # detect and wire every supported client

This registers the MCP stdio bridge (agentcache.mcp_stdio) and, with --with-hooks, installs workspace hook blocks that call /agentcache/observe on file events. Re-running is safe — the CLI detects stale entries (wrong interpreter path, missing env keys, etc.) and repairs them, or prints a --force hint when the entry is already up to date.

See docs/mcp-setup.md for per-client config paths, manual wiring, and troubleshooting.

Auth

Set AGENTCACHE_SECRET to require a shared secret on write routes:

export AGENTCACHE_SECRET=your-long-random-string
agentcache serve

Clients send it via Authorization: Bearer <secret> or ?secret=<secret>. Unset = open (fine for localhost; do not expose an unauthenticated instance).

Deployment

Dockerfile and docker-compose.yml are included. For multi-worker WSGI (Gunicorn + workers), run the server without background threads and use a dedicated worker process:

agentcache serve --no-workers                    # web tier
agentcache worker --tasks=index,forget           # sidecar

Details and the ongoing hardening plan live in docs/publishing_roadmap.md.

Compatibility notes

• Routes are dual-mounted at both /agentcache/* and /agentmemory/* for backward compatibility with the previous package name.
• agentcache.legacy is a thin re-export shim retained for callers that imported from the pre-split module layout. It will be removed in 1.0 — migrate imports to agentcache.core.* when convenient. See docs/adr/ for the split rationale.

Development

git clone https://github.com/Yashwant00CR7/agentcache
cd agentcache
pip install -e ".[dev,local-embeddings]"

pytest                     # test suite
ruff check .               # lint
ruff format .              # format
python -m build            # build sdist + wheel into dist/
twine check dist/*         # validate metadata before upload

Links

• Source: https://github.com/Yashwant00CR7/agentcache
• Issues: https://github.com/Yashwant00CR7/agentcache/issues
• License: MIT — see LICENSE