Skip to content

memex

Memex

A durable, local-first memory layer for AI coding agents.

The memory store is the filesystem: every memory is a human-readable Markdown page under ~/.memex/docs/, git-able and editable by hand. SQLite FTS5 is a disposable BM25 index — delete it and it rebuilds from the pages. No server, no cloud, no embeddings.

memex write --type preference --title "Deploy on Fridays" \
    --body "The team deploys to production on Fridays only."
memex recall "deploy"

Why three layers

Agents forget to call tools. Memex does not rely on them remembering:

  • Pull — eight typed MCP tools (memex serve-mcp) for model-initiated memory operations, with enums and bounds enforced in the tool schemas.
  • Push — harness hooks (memex hook session-start | prompt | transcript) inject relevant memories into context on every turn and capture session transcripts automatically. Deterministic, no model cooperation required.
  • Proof — memex verify turns memory practice into a CI gate: health checks always, recall/write evidence on demand, exit 1 fails the build.

One contract, every harness

Harness Push Pull Capture
pi extension: per-turn injection stdio MCP session JSONL
Claude Code SessionStart / UserPromptSubmit / SessionEnd hooks stdio MCP transcript
Codex AGENTS.md contract + notify stdio MCP rollout
Copilot CI carries it remote (future) verify workflow

Install any adapter with memex install <name> (the marketplace ships inside the package).

Provenance by construction

Every captured session becomes an episode node linked to its raw JSONL transcript. Any memory can be traced back to the conversation that produced it — direct when the node came from a transcript, inferred when an episode references it.

Next