Agent memory you can diff, blame, and roll back
Talewell gives every agent platform the same long-term memory — git-tracked Markdown, deterministic recall, and provenance gates — with no vector database in sight.

What it does
Talewell is a long-term memory layer for AI agents and the reference implementation of AMP (Agent Memory Protocol). Each memory is one Markdown file with frontmatter — id, kind, aliases, importance, links — and every write is a git commit, so the corpus is diffable, greppable, and rollback-able. Recall is deterministic: relevance × recency decay × importance, computed locally, no model call. It reaches hosts three ways — an MCP server exposing six memory tools, a native OpenClaw plugin that injects a compact index before every turn, and a plain file convention that writes a managed block into AGENTS.md or CLAUDE.md for anything that reads instruction files.
The interesting bit The README credits two ancestors: a ‘$2.5B AI assistant’ whose memory is reportedly git-tracked Markdown plus grep, and OpenClaw’s memory architecture. The unusual part is provenance as a hard gate — user-stated or agent-derived content gets stored, while web pages, tool output, and other people’s content are structurally refused. A security property, not a tuning knob. Writes also skip the reply path entirely, going through gated consolidation instead.
Key highlights
- Corrections supersede: the old value retires from recall and prompt injection while git keeps the history.
- Records cross-reference via
[[id]]links with stable IDs and aliases — a navigable link graph in plain files. - Fully offline baseline: no API key needed to search your own memory; embeddings are an optional layer, not the kernel.
- Verified on OpenClaw, Claude Code, Codex, OpenCode, and Pi; unverified hosts get safe defaults rather than guessed config paths.
- Zero npm dependencies in the kernel, with 28 tests covering the taint gate, supersession, rollback, and deterministic CJK recall.
Caveats
- It’s v0.1, and four hosts in the registry (Hermes, Grok bot, Muse AI, DeepSeek harness) are explicitly unverified.
- The ‘$2.5B assistant’ anecdote is unnamed and unverifiable — treat it as color, not evidence.
- The project grew out of an OpenClaw-specific skill and is ‘Powered by MyClaw.ai,’ a hosted OpenClaw service, so the platform-neutral pitch carries visible ecosystem gravity.
Verdict If you run agents across platforms and want memory you can audit in git and read as plain files, this is a rare bird. If your memory problem is really a semantic-search problem, the vector-free kernel will feel like a constraint — though embeddings remain an optional layer.
Frequently asked
- What is LeoYeAI/talewell?
- Talewell gives every agent platform the same long-term memory — git-tracked Markdown, deterministic recall, and provenance gates — with no vector database in sight.
- Is talewell open source?
- Yes — LeoYeAI/talewell is open source, released under the MIT license.
- What language is talewell written in?
- LeoYeAI/talewell is primarily written in JavaScript.
- How popular is talewell?
- LeoYeAI/talewell has 550 stars on GitHub.
- Where can I find talewell?
- LeoYeAI/talewell is on GitHub at https://github.com/LeoYeAI/talewell.