| name | agent-memory-palace |
| description | Give an AI agent persistent, navigable memory across sessions using a memory-palace structure. Use when building agent memory, when the user says "remember this across sessions", "give the agent long-term memory", "persistent context", "knowledge graph for my agent", or for coding agents, Claude cowork, and ChatGPT projects that must retain state. Reads and writes palace.json as a durable spatial knowledge store the agent reloads each session. |
agent-memory-palace
The agent-facing application of the method. Instead of memorizing for a human, the agent keeps
its own durable memory as a palace: rooms are domains, stations are topics, loci are facts —
a structure that is both human-legible and machine-navigable.
When to use this
An agent (coding assistant, cowork session, ChatGPT/Claude project) needs to retain knowledge
beyond a single context window: project conventions, decisions, user preferences, API contracts,
recurring facts. This skill defines how the agent stores and retrieves them.
Instructions
-
One palace file, known location. Keep a single palace.json (type: "memory-palace") at a
stable path the agent reloads at session start — e.g. .agent/memory/palace.json in the repo, or
the project's knowledge directory. It is the agent's durable memory.
-
Map the structure for knowledge, not lists:
room = a domain (e.g. "build system", "API contracts", "user preferences").station = a topic within the domain.locus = one fact, with target (the fact, terse), fact (elaboration), and optional refs
(file paths / URLs that ground it).
-
Write protocol (capture): when a durable fact appears (a decision, a convention, a fix), add
or update a locus. Keep target atomic and self-contained. Record refs so the fact is
verifiable, and recall.lastReviewed as the write timestamp. Append, don't silently overwrite —
if a fact changes, update it and note what changed.
-
Read protocol (session start): load the palace, skim room/station names to rebuild a map of
what is known, and pull the loci relevant to the current task. Treat it as memory to verify
against source, not as ground truth — facts can go stale (re-check refs).
-
Retrieve protocol (during work): match the task to the nearest room → station → loci. Prefer
structural navigation (domain → topic) over scanning everything; that is the efficiency the
palace buys an agent.
-
Keep it portable. Same schema as the human palace, so palace-visualizer can render the
agent's memory and a human can inspect it. Tolerate unknown fields; default missing ones.
-
Prune. When a room grows noisy, consolidate near-duplicate loci and archive obsolete facts
(move to an archive room rather than deleting outright).
Refusals
- Never store secrets, credentials, or private user data in the palace file.
- Never treat a stored fact as authoritative about current state without re-checking its
refs —
memory is a cache, the source is the truth.
- Don't let the file sprawl unstructured — every fact lives in a room and a station.
References
spec/MEMORY-PALACE-METHOD.md §5 (schema, agent mapping).spec/palace.schema.json — the contract this skill reads and writes.
Built on SIP · mind-palace-agent-skills · Memory Palace Method v0.1
