// the find
akitaonrails/ai-memory
Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
ai-memory gives AI coding agents (Claude Code, Codex, Cursor, etc.) a persistent wiki so context survives session ends and agent switches. It captures prompts and tool calls via lifecycle hooks, compiles them into Karpathy-style markdown pages at session boundaries, and prepends a handoff block when the next agent starts. Built in Rust, stores in SQLite + git-versioned markdown.
The 'compile not retrieve' approach is the right call — synthesized narrative pages beat raw log retrieval for both quality and token budget. Zero-LLM mode with FTS5 search still works, so you get value before spending money on an embedding provider. Plain markdown on disk means the wiki is inspectable, grep-able, and backed up with rsync — no proprietary blob storage to babysit. The multi-agent support matrix is genuinely broad and the install-mcp/install-hooks idempotent commands make wiring up a new agent a two-liner.
The README acknowledges that Grok ignores SessionStart stdout, so handoff injection silently breaks for that agent — you only recover via MCP after the fact, which defeats the main value proposition for Grok users. The anthropic-oauth mode (Claude subscription instead of API key) is documented with a warning that it violates Anthropic's usage policies, which is a liability for anyone who takes that path seriously. Cross-agent session continuity assumes the next agent runs in the same working directory; mono-repos or projects where engineers frequently work across multiple worktrees need marker files and careful configuration to avoid project-routing splits. At v0.8 with 24 SQLite migrations and a project that lists 'being built collaboratively with Claude Code' in the acknowledgements, the schema stability story is still unproven in production.