engram: Instructions file for Codex

Engram

Persistent memory for LLM agents, backed by an agent-memory zettelkasten vault. Six skills read from and write to the vault on demand: recall, learn, please, route, curate, and write-memory. Each skill ships in agent-instructions/skills/ via engram update. Additionally, engram update/engram register-skills (--accept/--decline/--dry-run/--adopt) offer a companion vault runbook note for every skill, command and Pi prompt template in the default folders — not only engram's six: Claude user skills and commands (symlinks followed), claude.ai-synced skills, Pi user and agents skills, Pi-configured skill/prompt paths and packages, enabled plugin skills and commands, and project-local Claude sources plus Pi sources for trusted projects. Each note is identified by a source-qualified skill key (claude:route, claude:cmd:audit, superpowers:brainstorming, commit:cmd:commit, project:github.com/toejough/engram:cmd:opsx:apply), with a slug derived from the key and skill_key/skill_source recorded on the note; every key is source-qualified; engram's own skills are keyed by the folder they are found in (e.g. claude:route, pi:route), identical copies collapse, and every preamble names the note's skill_source. Removal offers are limited to sources that were read successfully. Declines are remembered by hash, keyed by skill key, in vault-root skill-registrations.json; nothing is written without a terminal, and offers are answered by key, key-prefix pattern, or @scope. New/refreshed notes are pending offers the curate skill reviews. recall and learn invoke write-memory natively as a skill at their write sites, not reached by basename/wikilink. please is one note (its former three sub-runbooks merged into the single note/skill body). The runbook concept itself (captured runbooks with triggers, the shim follow-frame) remains unchanged — only the "skills are replaced by runbooks" direction is reversed.

Automatic memory, on Claude Code only, is a mod, not a skill: engram update deploys agent-instructions/mods/engram-memory, which runs engram auto deliver for each prompt (the vault notes that bear on it arrive beside it) and, after each turn, engram auto judge and engram auto write (each lesson of the turn becomes a note marked written automatically) and engram auto sweep (the session's transcript is indexed for engram query). What those commands did, or could not do, is recorded under auto in the engram data directory, and engram auto deliver prints it on standard error at the next prompt, as at most four lines (the notes delivered, written, left out and failed) and removes the records it showed; the mod shows each line through Claude Code's log ($.ui.log): the user sees them and the model does not read them. It acts only for the prompts the user sends (origin composer, bridge, slack-ping, sdk); turns of headless prompts (sdk, claude -p) are judged and swept only with ENGRAM_AUTO_CAPTURE_SDK=1.