using-memory

Using memxt memory

You have a persistent, local memory palace via the memory MCP server. It survives across sessions. Nothing leaves this machine. No cloud LLM is used to compress or store memory (unlike claude-mem-style pipelines).

Progressive disclosure (save tokens)

Always prefer this 2-step flow:

1. memory_search — default detail=index returns compact hits (#id, wing/room, source, ~100-char snippet). Cheap.
2. memory_get — pass promising id or ids to load full verbatim bodies only for what you need.

Do not pass detail=full unless the query is tiny or you already know you need every body (wastes context).

Example:

memory_search({ query: "cart limit", limit: 5 })
→ #42 decisions/… "Cart is capped at 37…"
memory_get({ ids: [42] })
→ full drawer text

Other tools

• memory_wake_up — continuity brief (also auto-injected at SessionStart).
• memory_profile — stable project facts; no embedding model.
• memory_store — persist a decision / constraint / fact (verbatim). Prefer room: "decisions" for architecture choices.
• memory_forget — delete by drawer id.
• memory_stats / memory_dream — health / consolidation.

When to recall

Before answering "what / why / how did we …", "last time", "remember when", "our convention for …", or anything that depends on history:

1. memory_search (index)
2. memory_get on relevant ids
3. Cite #id / source when you use a memory

When to store

After a real decision ("let's use X because Y"), a discovered constraint, a correction the user makes, or a fact you'd want next session, call memory_store. One crisp memory per fact.

Good:

"Cart is capped at 37 items because the Brightwell ERP rejects larger orders (0x5C error)."

Bad: "we talked about the cart."

Hooks already autosave session tails on Stop + PreCompact (local, verbatim). Still call memory_store for important decisions so they land in the profile, not only as a long episode.