Documentation / v3

Start with
memory.

Install CLM, save the first durable fact and connect every CLI agent to one local source of context.

01 / Quickstart

Install CLM.

The core package includes SQLite storage and FTS5 keyword recall. Add the optional local embedding model for semantic retrieval.

CLM writes its database to ~/.clm/clm.db. Override that location with CLM_DB_PATH.

02 / Capture

Save what changed.

Store durable facts, decisions, preferences and project state. Pin only context that should appear at every startup.

--kind

note, fact, preference, project, decision, bug, identity or task

--importance

Relevance weight from 0.0 to 1.0

--pin

Include in the automatic hot layer

03 / Retrieve

Ask naturally.

Recall combines exact search, semantic similarity and typed associative spreading. The output includes the score and retrieval path.

04 / Architecture

Two layers,
one database.

Hot

Automatic continuity

Pinned and salient memories are written between managed markers in an agent startup file.

Cold

On-demand recall

The full local store stays queryable through the CLI and MCP without filling every context window.

05 / Integrations

Connect your agents.

Claude Code

Codex

[mcp_servers.clm]
command = "python"
args = ["~/ai/clm/mcp_server.py"]

Any MCP-speaking client can use the same five tools: save, recall, link, boot and stats.

06 / Security

Memory is not
a secret store.

SecretGate rejects credential-shaped content at the write boundary. Store secrets in 1Password and save only their op://Vault/Item/field references in CLM.

Read the security model ↗