Wire the two forward-declared mempalace.memory_mode modes so they actually route recall/capture instead of silently behaving as `augment`: - kg_backend: the palace temporal KG is the primary knowledge-graph source; native .planning/graphs/ is the fallback. Non-KG drawer recall stays additive. - replace: recall resolves through the palace as the source of truth; native artifacts are the fallback. Every mode stays onError:skip and default-resilient — an unreachable palace degrades to native memory and GSD keeps writing .planning/graphs/, so no memory is lost. Cross-mode .planning/graphs/ migration remains a documented open question (PRD/ADR §17), out of scope here. Surfaces updated (instruction-only contract): recall/capture commands (+ generated skills), discuss/wave fragments, curator agent, capability.json schema. Docs: how-to Step 3, CONFIGURATION, FEATURES, CONTEXT glossary. Regenerated capability-registry, golden install-parity fixtures (mempalace hashes only), agent-size-baseline. Added a routing-contract + cross-surface parity test. Incidental (folded per no-defer rule): removed pre-existing unused imports (spawnSync in capability-registry.test.cjs; fs in issue-498-package-identity.test.cjs) that eslint flagged in/alongside the touched files. Closes #2007 Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
6.7 KiB
How to enable cross-session memory with MemPalace
Give GSD durable memory across sessions and projects. When MemPalace is connected, GSD recalls prior decisions and patterns before you plan, and captures phase artifacts verbatim at each phase boundary.
What you need: MemPalace installed locally. GSD never installs MemPalace for you — pip install mempalace is your action. A working MemPalace install is not required to complete GSD work; the capability is onError: skip at every hook, so an absent or unreachable MemPalace leaves the loop unchanged.
Step 1 — Install MemPalace
Follow the MemPalace installation guide. The quick path:
pip install mempalace
mempalace init # creates the local palace (ChromaDB + SQLite)
mempalace status # verify the server is reachable
For interactive use (Claude Code), MemPalace's MCP server must be running so mempalace_* tool calls resolve. For headless or cron runs, the CLI path (mempalace wake-up, mempalace search, etc.) is used automatically — no MCP server needed.
Step 2 — Enable the capability
Inside your GSD project:
gsd-tools query config-set mempalace.enabled true
That is the only required step — mempalace.enabled is the master switch that gates all loop hooks. All other mempalace.* keys are optional refinements of the enabled behavior; they are honored at runtime by the skills, curator, and fragments.
Step 3 — Choose a memory mode
The mempalace.memory_mode key controls how authoritative MemPalace is during recall and capture — how tightly it couples to GSD's native memory (.planning/graphs/, STATE, learnings). All three modes are wired. Every mode is onError: skip and default-resilient: an unreachable palace degrades to native memory, and GSD keeps writing .planning/graphs/, so no mode risks memory loss.
| Mode | What it does | When to use it |
|---|---|---|
augment (default) |
The palace is an additional recall layer alongside .planning/graphs/ and learnings. Lowest coupling — native memory stays authoritative and the palace supplements it. |
Most users. Safe to enable immediately. |
kg_backend |
Knowledge-graph queries resolve against the palace's temporal KG as the primary source; .planning/graphs/ becomes the fallback. Non-KG drawer recall stays additive. |
You want the palace's temporal KG to drive decision recall while keeping native graphs as a safety net. |
replace |
Recall resolves through the palace as the source of truth; native artifacts are consulted only as a fallback when the palace is unreachable. | You want the palace to be the authoritative memory store for this project. |
Set the mode with:
# augment is the default; most users need no change.
gsd-tools query config-set mempalace.memory_mode kg_backend # or: replace
Note — cross-mode migration. Switching an established project to
kg_backendorreplacechanges how new recall and capture resolve; it does not retro-migrate memory already written to.planning/graphs/into the palace. Backfilling existing native memory into the palace is a separate, not-yet-implemented concern — choose the mode you want at the start of a project for the cleanest result.
Step 4 — Run a phase and observe recall and capture
With mempalace.enabled: true, here is what you will see at each loop point:
Discuss phase (/gsd-discuss-phase):
- At
discuss:pre— MemPalace recall fires: prior decisions, patterns, and surprises are surfaced as context before the discussion begins. - At
discuss:post—CONTEXT.mdis filed into thedecisionsroom in MemPalace.
Plan phase (/gsd-plan-phase):
3. At plan:pre — a MEMORY-RECALL.md file appears in the phase directory containing prior decisions, patterns, and surprises retrieved from the palace.
4. At plan:post — PLAN.md is filed into the planning room.
Verify phase (/gsd-verify-work):
5. At verify:post — SUMMARY.md is filed into milestones in MemPalace.
Ship phase (/gsd-ship):
6. At ship:post — the gsd-mempalace-curator agent writes a diary entry and (if enabled) proposes cross-project tunnels.
Execute phase (/gsd-execute-phase, each wave):
7. At execute:wave:post — confirmed problem→fix pairs are captured into the problems room via the capture-problems fragment.
If MemPalace is unreachable at any step, a skip-notice is written and the loop continues normally.
Optional configuration
Turn off recall or capture independently
# Disable recall injection at discuss:pre (keeps capture)
gsd-tools query config-set mempalace.recall_on_discuss false
# Disable the MEMORY-RECALL.md step at plan:pre
gsd-tools query config-set mempalace.recall_on_plan false
# Disable artifact capture at phase boundaries
gsd-tools query config-set mempalace.capture_artifacts false
Cross-project tunnels
Enable tunnel proposals at ship:post to connect related rooms across projects:
gsd-tools query config-set mempalace.cross_project_tunnels true
The curator agent will call mempalace_find_tunnels and propose connections to the wings it finds semantically related to this project's wing.
Passive mid-session capture (reserved — not yet implemented)
mempalace.auto_capture_hooks is a forward-declared key reserved for a future "Connected Capability" phase. Setting it to true currently has no effect — no native Claude Code hooks (stop, precompact, session-start) are installed by this key yet. The capability's hooks array is empty; the deliberate loop hooks are the only active integration today.
# Not yet functional — reserved for a future release
# gsd-tools query config-set mempalace.auto_capture_hooks true
Override the wing name
By default, the wing name derives from project_code or the project directory. Override it:
gsd-tools query config-set mempalace.wing my-project-name
What to expect when MemPalace is absent
If MemPalace is not installed, not running, or unreachable:
- Every hook logs a skip notice and continues.
MEMORY-RECALL.mdis written with an "unavailable" stub (the planner can still proceed without it).- No phase step fails or blocks.
- Loop behaviour is identical to having
mempalace.enabled: false.
You can safely leave mempalace.enabled: true in a config that will be used on machines without MemPalace — it is safe to do so.
Full configuration reference
See MemPalace Settings in the Configuration Reference for the complete key list with types and defaults.