memfmt: A Plain-File Agent Memory Format That Tracks Success and Failure Counts
No_Advertising2536 · reddit · 2026-09-04
A developer released memfmt, a dependency-free plain-file memory format for AI agents, fixing two flaws in existing markdown-memory tools (EverOS, basic-memory, etc.): incompatible layouts, and storing only what agents should do — never whether it worked.
- Procedures carry success/failure stats: e.g. 11 ✓ / 1 ✗, smoothed to 86% reliability via a prior so one success never reads 100%; agents can distinguish battle-tested workflows from stale notes
- Evolution logs inside each procedure explain why a version changed
- Entities/episodes/procedures split, with [[wikilinks]] for Obsidian graphs and git giving diff, review, revert over what your agent learned
- CLI: memfmt stat/validate/context; validate round-trips files to guarantee no tool rewrite loses data
- Honest limits: word-overlap relevance (no embeddings), multi-machine sync unsolved; author sells a hosted memory product but the library is deliberately unbound to it
More from coding & agent
- Open source coding agent opencode passes 200k GitHub stars — anomalyco · 2026-09-04
- Ian Nuttall burns through all Ahrefs MCP credits in one week, falls back to computer use — iannuttall · 2026-09-04
- Open-source autoresearch turns coding agents into autonomous improvement loops (6k stars) — tom_doerr · 2026-09-04
- First-ever AGNTCon + MCPCon world tour kicks off: Shanghai, Tokyo, Amsterdam — marlene_zw · 2026-09-04
- 5 embedding compression techniques explained: 10M vectors shrink from 62GB to 2GB — blaizedsouza · 2026-09-04
- WASM + WebGL texture flag trick enables fast per-pixel updates in browser — YishengJiang · 2026-09-04