agent-memory-hub

A local-first, human-auditable memory system for personal AI agents (Codex, Claude Code, and any MCP client). One canonical memory you own, queried by every agent through MCP — instead of each tool keeping its own siloed, opaque memory.

Status: v0.1 MVP. The recall + governed-write loop works end to end, with CI, a dedup/conflict hint at review, staleness hygiene, and session checkpoints. Heavier designed capabilities (full auto-classification, hybrid/vector retrieval, daily insight, multi-device migration, …) are on the ROADMAP, not built yet. This repo is honest about that line.

Why

• You own it. Plain Markdown in Git — readable, diffable, revertable, portable.
• Every agent shares one memory. MCP exposes search/recall/propose to any client.
• Writes are governed. Agents propose; a human approves. Untrusted content

can never auto-promote to a rule.

• Secrets never live in the vault — only a macOS Keychain reference.

Core loop

• Recall at task start: memory.context_pack (task-scoped, policy-filtered,

active-only), plus memory.search / memory.get / memory.answer.

• Archive at task end (the Durable Learning Checkpoint): the agent surfaces

1–3 candidate learnings; only after explicit user confirmation it calls memory.propose_write → inbox/proposals/ (status proposed, never canonical).

• Approve (human): agent-memory-cli proposals approve promotes it into