Parent: #164
Goal
Add a durable project-memory store for semantic decisions, rejected paths, episodic lessons, and procedural skills, with provenance and supersession.
This is not a transcript archive. It is the curated memory layer that lets the maestro resume with load-bearing facts and old rejected approaches without carrying low-level worker context forever.
Scope
Expected owned surface:
src/forge_loop/memory/models.pysrc/forge_loop/memory/curator.py- a new memory store module under
src/forge_loop/memory/;
- focused tests under
tests/test_memory_store.py or equivalent.
Keep the first implementation small and explicit. SQLite is acceptable if it follows the existing eventlog style; YAML/JSON is acceptable only if the tests cover durable reopen and supersession correctly.
Required Behavior
- Persist memory items with:
memory_id;- kind: semantic, episodic, procedural;
- title/body;
- tags;
- provenance/source event or source task reference;
- active/superseded state;
- created timestamp.
- Query active memory items.
- Query rejected-path memory independently.
- Supersede an item and exclude superseded records from active boot context.
- Preserve supersession provenance so old rationale is still inspectable.
- Keep the curator strict: only promote candidates with a real reason to remember.
Acceptance Tests
- Active semantic/episodic/procedural items round-trip after reopening the store.
- Superseding a memory item removes it from active results but leaves it inspectable.
- Rejected-path query returns only rejected-path tagged/typed items.
- Invalid confidence/provenance is rejected using the existing model constraints.
- Curator promotion policy does not promote empty "reason to remember" candidates.
Non-goals
- Do not save raw transcripts.
- Do not implement embedding search yet.
- Do not generate summaries with an LLM in this ticket.
- Do not make memory promotion automatic from every worker event.
Verification
Run at minimum:
- New memory store tests.
env -u VIRTUAL_ENV uv run --extra dev pytest tests/test_eventlog_sqlite.py -q if using event refs.env -u VIRTUAL_ENV uv run --extra dev ruff check <changed files>env -u VIRTUAL_ENV uv run --extra dev ruff format --check <changed files>
Customer Story
A maestro resuming forge-loop after context loss benefits because load-bearing decisions and rejected approaches survive outside the model context.
Source
Expanded during Forge self-dogfood sprint planning on 2026-06-02.
Parent: #164
Goal
Add a durable project-memory store for semantic decisions, rejected paths, episodic lessons, and procedural skills, with provenance and supersession.
This is not a transcript archive. It is the curated memory layer that lets the maestro resume with load-bearing facts and old rejected approaches without carrying low-level worker context forever.
Scope
Expected owned surface:
src/forge_loop/memory/models.pysrc/forge_loop/memory/curator.pysrc/forge_loop/memory/;tests/test_memory_store.pyor equivalent.Keep the first implementation small and explicit. SQLite is acceptable if it follows the existing eventlog style; YAML/JSON is acceptable only if the tests cover durable reopen and supersession correctly.
Required Behavior
memory_id;Acceptance Tests
Non-goals
Verification
Run at minimum:
env -u VIRTUAL_ENV uv run --extra dev pytest tests/test_eventlog_sqlite.py -qif using event refs.env -u VIRTUAL_ENV uv run --extra dev ruff check <changed files>env -u VIRTUAL_ENV uv run --extra dev ruff format --check <changed files>Customer Story
A maestro resuming forge-loop after context loss benefits because load-bearing decisions and rejected approaches survive outside the model context.
Source
Expanded during Forge self-dogfood sprint planning on 2026-06-02.