project-memorylisted

# Project Memory Durable memory that survives across sessions, so the agent stops relearning the same project facts. Three layers, cheapest first. ## The layers - **Decision memory = the ledger** (`references/core/ledger.md`). Every elected truth is already durable, append-only, and carries a `flip_criteria` (when to reopen). Do NOT duplicate decisions here — point at the ledger. - **Project memory = `MEMORY.md`** at the repo root: a short, human-readable, git-tracked list of durable facts the ledger doesn't hold — conventions ("we use pattern X"), gotchas ("Y looks wrong but is intentional"), environment quirks, and user preferences. Loaded as always-on context via `AGENTS.md`; edited deliberately, never a dumping ground. - **Graph memory (optional) = the `cognee` MCP** (`cognee/cognee-mcp`): a queryable, self-editing knowledge graph for larger, associative recall when `MEMORY.md` is not enough. It supports **deliberate writes** (`cognee.remember("…")`) — use those, not conversational auto-capture, so this stays curated and not a dump. **It is not wired for you, on purpose.** Unlike the other servers this package declares, cognee runs its own LLM extraction: it needs a Docker container and an `LLM_API_KEY`. Declaring it by default would hand every user a server that fails to connect — so it is opt-in, and the first two layers cover durable memory without it. To turn it on, start the container and add the server to your own MCP config: ``` do