Most agent memory lives in an opaque vector store nobody can open, audit, or fix. ReMe flips that premise: memory is a folder of plain Markdown files with frontmatter and wikilinks that both you and the agent read and edit directly. The payoff is that debugging what an agent "remembers" becomes reading and diffing a directory you can version in git.
What Sets It Apart
- Memory as file: each memory node is a Markdown file with frontmatter and wikilinks, so recall is inspectable and hand-editable rather than a black-box similarity score.
- Self-evolving pipeline: Auto Memory, Auto Resource, and Auto Dream progressively turn raw sessions and resources into daily notes and then long-term digests, wiring wikilink relationships automatically instead of leaving you to curate them.
- Progressive hybrid search: it blends wikilink graph traversal, BM25 keyword matching, and embeddings, so retrieval catches exact terms, semantic paraphrases, and related-by-link context in one pass — not just nearest-vector hits.
- Agent-native plumbing: it runs as a local service exposed over CLI and MCP, with a shipped Claude Code plugin and a QwenPaw in-process SDK path, so wiring
auto_memoryand search into an agent's lifecycle is a config step, not a rewrite.
Who It's For
Great fit if you want agent memory you can audit, hand-edit, and version in git — coding assistants that must keep project context across sessions, personal assistants, or traceable knowledge QA. The design is also backed by an ACL 2026 Findings paper on procedural memory, so the consolidation model isn't ad hoc.
Look elsewhere if you need a turnkey hosted memory API or a language-agnostic drop-in library: ReMe is Python 3.11+ and runs as a standalone service, and the file-backed model trades some disk and I/O overhead for its readability and durability.
