Summary
mcp_server/handlers/seed_project.py:180 purges memories globally by tag, regardless of the domain argument:
purged = 0
if not dry_run:
purged = _get_store().delete_memories_by_tag("seeded")This means every call to seed_project wipes all previously-seeded memory rows across every domain — only the most recent seed survives.
Reproducer
Sequentially seed two repos with distinct domains:
seed_project(directory="/path/to/repoA", domain="repo-a")
# stored 4, purged_stale 0
seed_project(directory="/path/to/repoB", domain="repo-b")
# stored 7, purged_stale 4 <-- wiped repoA's rows, despite different domain
memory_stats()
# only repoB's 7 rows + any pre-existing non-"seeded" memories
Expected: purged_stale 0 on the second call, since repo-b has no prior seeded memories.
Actual: every prior seeded-tagged memory is deleted regardless of domain.
Impact
- Bootstrapping multiple repos in one session (the natural workflow on a multi-project dev machine) leaves only the last repo's structural-summary / configs / docs / entry-points in memory.
- Knowledge-graph entities and relationships accumulate fine across calls (different table, different lifecycle), so cross-repo recall via the KG still works. The loss is the per-repo discovery memory rows.
- Workaround: end with the repo whose rows you most want to retain.
Suggested fix
Scope the purge by domain when one is provided:
purged = 0
if not dry_run:
purged = _get_store().delete_memories_by_tag_and_domain("seeded", domain) if domain \
else _get_store().delete_memories_by_tag("seeded")
Or keep the global purge but filter by directory (since each seed has a directory field on the stored memory).
Environment
- Cortex 3.14.5, Docker install
- PostgreSQL 15 + pgvector
- Discovered while bootstrapping 8 sibling repos with
seed_project per repo
Notes from same session (separate, lower priority)
While verifying the above, also surfaced:
remember, recall, and get_telemetry return structured_content must be a dict or None. Got str: ... from the FastMCP layer despite the underlying op succeeding. Likely missing/incorrect output_schema declaration on those handlers.query_methodology(cwd="C:/Users/...") does not slug-normalize Windows-style paths to the existing domain; passing project= works.
Happy to file these as separate issues if useful.
Summary
mcp_server/handlers/seed_project.py:180purges memories globally by tag, regardless of thedomainargument:This means every call to
seed_projectwipes all previously-seeded memory rows across every domain — only the most recent seed survives.Reproducer
Sequentially seed two repos with distinct domains:
Expected:
purged_stale 0on the second call, sincerepo-bhas no prior seeded memories.Actual: every prior
seeded-tagged memory is deleted regardless of domain.Impact
Suggested fix
Scope the purge by domain when one is provided:
Or keep the global purge but filter by
directory(since each seed has adirectoryfield on the stored memory).Environment
seed_projectper repoNotes from same session (separate, lower priority)
While verifying the above, also surfaced:
remember,recall, andget_telemetryreturnstructured_content must be a dict or None. Got str: ...from the FastMCP layer despite the underlying op succeeding. Likely missing/incorrectoutput_schemadeclaration on those handlers.query_methodology(cwd="C:/Users/...")does not slug-normalize Windows-style paths to the existing domain; passingproject=works.Happy to file these as separate issues if useful.