Local system · memory online

Agents forget.
CLM remembers.

A local-first memory layer that gives every CLI agent continuity—across sessions, tools and the work that matters.

Install CLM ↓See how it works
01 / 05
Persistent context
without the cloud.
The continuity problem

Your agent can reason for hours, make decisions, learn your preferences—then begin the next session knowing none of it.

CLM turns finished sessions into durable context.
Two layers. One memory.

Present when needed.
Invisible when not.

Important context arrives automatically. Everything else stays searchable, local and out of the way.

Read the architecture →
01
Hot layer

Continuity at startup.

Pinned facts and high-salience memories are injected into each agent's startup context. No prompt. No tool call. It simply begins informed.

clm boot ~/.claude/CLAUDE.md
02
Cold layer

Recall on demand.

Keyword search, local embeddings and associative links find the right memory—even when your question uses different words.

clm recall "how do we ship?"
Hybrid retrieval

Meaning,
not matching.

01

Exact signal

FTS5 and BM25 find precise language instantly.

02

Semantic signal

Local embeddings understand paraphrase and intent.

03

Associated signal

Typed links surface decisions connected through real work.

Ask naturally

The words can change.
The memory stays.

Try the local demo. CLM ranks direct and associated memories, explains why they surfaced, and never sends your context to an API.

CLM / semantic recall local / ready
›
Try:
#42preferencescore 0.73

Deploy commercial projects by git push; Vercel handles the release.

#11identityscore 0.71

Scott is a solo operator, builder and trader who prefers direct, action-first replies.

#18decisionscore 0.67

Keep production releases reversible and verify the live path after every deploy.

3 memories surfacedlocal simulation · no network
100%local
Built for trust

Your memory
stays yours.

One inspectable SQLite database. Offline embeddings. WAL-safe access for multiple agents. SecretGate blocks credentials before they enter memory, while 1Password references keep secure operations useful.

  • 01No cloud account
  • 02No telemetry
  • 03No vendor lock-in
  • 04Human-readable records
Shared context forClaude CodeCodexKimiAny MCP client
Open source · MIT

Give your agents
a past.

FTS5 keyword recall and SQLite storage

Python 3.10+ · macOS and Linux · fully local at runtime

Quickstart documentation →Read the source on GitHub ↗