← ClaudeAtlas

project-memorylisted

Use the project's engram memory to stay aligned with the codebase — recall past decisions, architecture and constraints, and record durable new facts. Invoke this whenever you need project context you don't already have, or when a decision/constraint worth remembering emerges. Backed by the `mem` CLI (zero-token deterministic core); no API key.
max423/engram-memory · ★ 1 · AI & Automation · score 77
Install: claude install-skill max423/engram-memory
# Project memory (engram) This project keeps a curated, git-native memory under `.memory/`. Use it instead of scanning the whole repo. The `mem` CLI is the interface (if not on PATH, use `python3 core/mem.py`). Everything here is token-minimal — **never read the whole wiki**. ## When to READ from memory (do this proactively) Before exploring the codebase or answering an architecture/decision question: 1. Get the big picture: read `.memory/context.md` (the code-aligned map) — or run `mem digest` for map + decisions catalogue in one shot. 2. Narrow deterministically (0 tokens): `mem search "<key terms>" --top 5` (`--type decision`, `--tag <t>`, `--backlinks <slug>` to focus). 3. Open ONLY the 1–4 pages the search surfaces. Check their `sources:` if you need to verify a claim against the raw source. Cite pages by slug, e.g. `[[reconcile-al-merge]]`. If memory doesn't cover it, say so — don't invent. ## When to WRITE to memory (record durable facts autonomously) When a **durable** fact emerges during the work — a decision made, a constraint discovered, "module X is responsible for Y", a tradeoff chosen — record it: ```bash mem note "Postgres chosen over Mongo: we need transactions and team knows it" ``` `mem note` writes a `raw/` source (the source of truth, anti-drift intact) and compiles it into a page. Record the **why**, compactly. Do NOT record trivia, one-off lookups, or things already in memory (search first). Prefer one atomic fact per note. ## Keepin