engramory

When building memory for AI agents, the two usual traps are: turning memory into a heavyweight database/vector-retrieval system, or letting a memory folder grow without bound until it becomes a pile of junk nobody dares to touch. Engramory takes a third path — zero infrastructure, no embeddings, no server. Just a folder of small, human-readable Markdown files plus an index that loads every session, and that turns out to be enough to keep memory in order.

The core is a strict but practical discipline: dedup before writing, update instead of duplicating, delete when wrong, and never store what the repo already records. Every feedback note must state its Why and How to apply, so procedural memory is actually actionable. The index is deliberately bounded — since the loading window is finite, it warns you to compact before old memories silently fall off the end of context. These rules load as standing rules, so they fire on every task rather than depending on a skill being retrieved by chance.

Who is this for? If you use Claude Code, Codex, or any agent host that supports standing rules and file reads, and you want a lightweight, auditable, hand-editable memory setup — or you are tired of auto-extracted vector fragments you cannot control — Engramory is for you. It does not try to out-search, out-scale, or out-automate the big memory systems. It bets on hand curation, a tiny always-loaded index, and a discipline you can carry to any agent, with everything inspectable in a plain text editor.