Personal Productivity

How LeoYeAI Built Talewell, Git-Backed Memory for AI Agents

Built by LeoYeAI · · Source: GitHub

Quick answer

LeoYeAI built Talewell, an open source long-term memory layer for AI agents that stores each memory as a Markdown file in a git repo. It replaces the earlier OpenClaw Auto-Dream skill and works with OpenClaw, Claude Code, Codex and other hosts.

Result: No usage results stated. The repo page shows 550 GitHub stars and 28 forks, and the project is at v0.1 with 28 tests.

What they built

LeoYeAI built Talewell, an open source long-term memory layer for AI agents. The README describes it as plugin-first memory for agent platforms and the reference implementation of AMP (Agent Memory Protocol). It is git-backed, uses plain Markdown, and needs no vector database.

The repo used to be OpenClaw Auto-Dream, a prompt-based memory skill for OpenClaw. According to the README, Talewell is the successor: the same idea, rebuilt as a portable plugin framework. The original skill is kept in a legacy-skill folder and the README says it still installs and works on OpenClaw.

The README says the design was inspired by a $2.5B AI assistant whose memory layer is git-tracked Markdown plus grep, and by OpenClaw's memory architecture. The project is sponsored by MyClaw.ai, a hosting service for OpenClaw.

How it works

  • One record per file. Each memory is a Markdown file with frontmatter fields like id, kind, aliases, origin, importance and links. Every write is a git commit.
  • Recall without a model call. Records are ranked by relevance, recency decay and importance, computed locally. The README says the baseline runs offline with no API key, and embeddings are an optional layer.
  • Provenance rules. Content marked as coming from untrusted or system origins is refused, so nothing gets written to memory without a trusted source.
  • Corrections and forgetting. Correcting a record retires the old value from recall, and forgetting removes a record. Git keeps the history in both cases.
  • Consolidation, not chat-time writes. Writes go through a gated, bulk consolidate step and stay out of the reply path.
  • Links. Records point to each other with [[id]] references, which forms a graph in plain files.

There are three ways to connect it. An MCP server exposes six memory tools. An OpenClaw plugin adds a compact memory index to the prompt before every turn. A file convention installer writes a managed block into AGENTS.md or CLAUDE.md. The README lists six verified hosts (OpenClaw, Claude Code, Codex, OpenCode, Pi Coding Agent and Hermes) and three unverified ones (Grok bot, Muse AI and a DeepSeek harness).

The current README does not describe a scheduled, sleep-style consolidation cycle for Talewell.

The result

The README makes no claims about usage results. The repo page shows 550 stars and 28 forks. The project is labeled v0.1, ships with 28 tests, and is MIT licensed.

How to build something similar

  • Start with plain files. Keep one fact per Markdown file in a git repo so you can read, diff and roll back what your agent remembers.
  • Label where each memory came from, and refuse writes from sources you don't trust, such as web pages or tool output.
  • Separate writing from chatting. Have the agent propose memories, then save them in a review or batch step.
  • Try simple ranking first (keyword match, recency, importance) and add embeddings only if recall is not good enough.

The original post

LeoYeAI · View on GitHub

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases