agent-memory

a memory system for AI coding agents. works with any harness — amp, cursor, codex, or manual CLI.

the problem it solves: your agent "forgets" everything between threads. you could grep a folder of notes, but that requires knowing the right keyword. nothing is always-in-context. everything is a flat bag of files with no navigational signal.

agent-memory fixes this with a four-layer pipeline:

the result: your agent starts each thread with relevant context already in AGENTS.md. no grep required.

how it works

• capture: your agent harness drops a journal entry into inbox/ after each session
• consolidate: an LLM reflects on the session, extracts knowledge, writes to topics/
• defrag: periodically reorganizes memory — merges duplicates, splits overgrown entries, assigns hot/warm tiers
• disclose: generates AGENTS.md with hot memory inlined, warm memory listed, cold memory omitted

pattern language

the system is built on 13 patterns extracted from axi-agent. see docs/pattern-language/ for the full treatment.

• WORKFLOW = DATA: a workflow is a record with state, not a running process. container restarts don't lose work — records rehydrate on boot.
• INJECT AT THE BOUNDARY: all I/O via adapter interfaces. machines call adapters; adapters call external systems. testable in isolation.
• EXHAUSTIVE STATE MODELING: every state and transition explicitly modeled. impossible states are compile errors.