02 · Overview
Three agents, one memory.
Claude Code, Codex and Cursor write into the same store and read from the same small core.
03 / 03
AI infrastructure · Tooling
AgentMemory
A shared long-term memory for AI coding agents: every session archived, distilled every fifteen minutes and folded nightly into a typed, sourced wiki.
Fig. 0 · Knowledge graph
Live
01
Problem
Coding agents start every session from zero. Context windows are large but finite, and filling them with history makes answers worse, not better.
Two requirements pull in opposite directions: nothing said should ever be lost, and what a session loads at start should stay small, fixed and trustworthy.
Fig. 1 · Pipeline
02.1 · Capture
Nothing gets lost.
Transcripts are read straight from disk, tracked by inode and byte offset, into an append-only archive. No fragile hooks: if a session crashes, its words are still there.
02.2 · Distill
Every 15 minutes, a log.
A small model turns new conversation into a dated log of decisions, facts and rule candidates. A systemd timer runs it; another commits the whole store every five minutes.
02.3 · Consolidate
Every night, a wiki.
A larger model folds the log into typed pages (decisions, entities, concepts, sources), each with provenance. Contradictions are flagged instead of silently overwritten.
02.4 · Boundaries
Small memory, hard walls.
Only a hard-capped core loads at session start, so context never bloats. Background workers run restricted and cannot touch human-owned notes without approval.
03 · In numbers
3Agents sharing one memory
15minDistill cycle
24hConsolidation cycle
04
Limits
- 01
A small model's summaries miss nuance. The raw archive is kept so nothing depends on them alone.
- 02
Contradictions are surfaced, not resolved: a human still decides which fact wins.
- 03
The core's size cap is a trade-off: what does not fit must be looked up, which costs a step.