Files
Claude-Code-Game-Studios/.claude/rules/agent-memory.md
T
DonchitosandClaude Opus 5.5 be8993bbc5 Release 1.1.3 — long skills keep their later steps
Fix release for 1.1.2: the 14 longest skills keep their later phases
after a compaction, a question nobody answered is never taken as an
answer, settings read the same from a subfolder, the commit check works
for a project in a subfolder of a larger repository, every skill opens
with a title, and the shipped allow rules no longer approve deleting a
branch or writing files without asking. Details in CHANGELOG.md; upgrade
steps in UPGRADING.md (v1.1.2 → v1.1.3).

Refs #83
Refs #133

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0185en4t76fPn7315q4c3nko
2026-10-08 17:52:48 +11:00

66 lines
3.0 KiB
Markdown

---
paths:
- ".claude/agent-memory/**"
---
# Agent Memory Rules
**Agent memory is one of three writes that need no "May I write?"** The other two
are named in `CLAUDE.md`:
- **Appends to `production/session-state/active.md`** — the session checkpoint
the hooks and the next session rely on. An append, never a rewrite of what is
already there.
- **A subagent's new file at the path its orchestrating skill named.** Only a
new file under `production/`, `docs/` or `tests/`, only at a path the skill's
prompt names (never one you chose), and only inside a phase the user approved.
`.claude/agent-memory/` is gitignored, per-agent, and never ships. Its exemption
exists because the memory is the agent's own working notes about how to do its
job, not a project artifact.
None of the three extends anywhere else. The Collaboration Protocol still governs
every other write under the code root (`src/`, `Assets/` or `Source/`),
`design/`, `docs/`, `production/` and `assets/`: ask first, naming the path.
Writing a memory is never a substitute for asking, and a memory must never be
used to record something the user declined to have written.
## Never record a temporary absence as a durable fact
This is the failure this file exists to prevent, and this repo is unusually
exposed to it.
**CCGS ships as a template the user clones *as their game*.** So on day one every
agent observes the same things — no game code, no GDDs, no engine configured, no
`tests/performance/` — and every one of those observations is **guaranteed to
stop being true**. A memory reading *"this repo is the framework, not a game;
perf requests have no target"* is accurate when written and actively harmful two
weeks later, when it tells a future run to return BLOCKED on a project that now
has a build. The agent will not re-derive it; that is what memory is for.
So:
- **Prefer recording a method over a state.** "Confirm an engine project and a
captured profile exist before profiling; return BLOCKED naming what is missing
rather than estimating" is durable. "There is no game here" is a timestamp.
- **If you must record a state, state what invalidates it**, on the same line:
`INVALIDATED WHEN: project.yaml declares an engine, or the code root (src/, Assets/ or Source/) contains game code.`
A reader with no other context must be able to tell whether the note still
holds.
- **Never record absence of a game, of assets, of tests or of config as a
settled property of the project.** Those are the states CCGS exists to move a
user out of.
## Disclose the write
Say in your response that you recorded a memory and where. An unmentioned write
is indistinguishable from no write, and the user cannot correct a note they do
not know exists — the same reason a skipped check has to announce itself
(`.claude/rules/skill-authoring.md`, obligation 3).
## Verify before you trust
A recalled memory reflects what was true when it was written. If it names a file,
a path, a setting or a count, **check that it still holds before acting on it**.
Memory is a starting point for investigation, never evidence.