Repository navigation
feat(memory): import Claude Code auto-memory into AutoBot knowledge base #16642
Description
Activity
- added a commit that references this issue
on Sep 13, 2026 - added a commit that references this issue
on Sep 13, 2026 Progress update — implementation is in PR #16643, currently parked.
Status: blocked, not a stall. Three security review rounds found real issues (all fixed except one pre-existing gap independent of this feature) plus two CI-mechanics violations (also fixed). Current head:
4d677c399.Acceptance criteria status:
- Importer parses
<memory-dir>/*.mdfrontmatter + body, configurable viaAUTOBOT_CLAUDE_MEMORY_DIRenv (never hardcoded) —knowledge/claude_memory_importer.py - Each memory file → one
knowledge_factsentry taggedcategory="claude_code_memory"+metadata.type - Writes go exclusively through the existing
knowledge.factswrite path (store_fact/update_fact) — no bypass ofstore_authority.py's declared system-of-record - Idempotent — deterministic
fact_id_for(slug), re-run updates rather than duplicates - Admin-gated (
check_admin_permission, router-level) —api/knowledge_claude_memory.py - Unit tests — 25 tests: frontmatter parsing, idempotent upsert, redaction (7 categories), owner/visibility metadata + enforcement proof
- Not yet met: "fact count before/after against a running instance, plus one imported fact retrievable via KB search" — requires a live host run, which requires merging first
Blocked on #16654 (canonical KB search + grounded-agent RAG enforce no ownership/visibility at all — pre-existing gap, independent of this feature, found during review; every PRIVATE-scoped fact in the system is exposed the same way today).
#16507was the other blocker; already resolved (merged via #16508).Non-blocking follow-ups filed during review, not required for this issue's own ACs: #16655 (exact-match fleet-hostname/IP redaction, closes a disclosed residual gap), #16656 (isolated knowledge buckets — better long-term home, not required now), #16661 (folder-watch continuous auto-import, an enhancement on top of the on-demand endpoint this issue asked for).
Full detail, evidence, and the review thread: PR #16643.
- Importer parses
- added a commit that references this issue
on Sep 14, 2026 - added a commit that references this issue
on Sep 19, 2026 Closure check against
origin/main(1edec3064c): 6 of 7. The last one is host evidence, and it blocks #16661Criterion Evidence Importer parses <memory-dir>/*.mdfrontmatter + body; path configurable, never hardcodedautobot-backend/knowledge/claude_memory_importer.py:parse_memory_file(:96),iter_memory_files(:123, excludesMEMORY.md).get_claude_memory_dir(:45-58) readsAUTOBOT_CLAUDE_MEMORY_DIR, else derives the path fromproject_root()andPath.home()with no machine path baked in. The comment at:38-41records why it is not inssot_config(frozen size ceiling).One knowledge_factsentry per file, tagged withmetadata.typeand a source markerCATEGORY = "claude_code_memory"(:36);_fact_metadatasetscategoryandmemory_type(:209-212)Writes only through the declared knowledge.factspathmodule docstring ( :11-13): upserts viastore_fact/update_fact, the pathstore_authority.pydeclaresIdempotent, keyed by slug fact_id_for(slug, owner_id)(:138);test_import_memory_file_updates_when_fact_present,test_import_claude_memory_is_idempotent_on_rerun(claude_memory_importer_test.py:263,:386)Admin-gated api/knowledge_claude_memory.py:35dependencies=[Depends(check_admin_permission)]; there is no caller-supplied path (:49-50)Unit tests: frontmatter parsing, idempotent upsert claude_memory_importer_test.py:test_parse_memory_file_*(:149-183), the idempotency tests above, plus owner scoping and redaction (:228-535)Evidence: fact count before/after on a running instance, and one imported fact retrievable via KB search Not met. No such evidence is recorded in this issue or its PRs. It needs an admin-authenticated POST /import_claude_memoryagainst a deployed instance, then a KB search. I did not run it; it requires admin credentials, which I do not read.Why this one matters more than its tick count suggests: this issue
blocks#16661 (continuous auto-import), so #16661 cannot be picked while one host observation is outstanding. Whoever can run an admin call against the deployed instance can close both the observation and the edge.What discharges the evidence criterion, and two things that would make the obvious run wrong
On the live instance, the importer's default directory does not exist for the backend service user, and
AUTOBOT_CLAUDE_MEMORY_DIRis not in the unit'sEnvironment=(EnvironmentFile contents not read).get_claude_memory_dir()falls back to~/.claude/projects/<project_root-with-/-as-->/memory(claude_memory_importer.py:45-58), which for a deployed checkout is a path nothing creates. A barePOST /import_claude_memoryshould therefore import zero facts, and zero would read as a failure of the importer rather than a missing directory.The second thing is #16654. It is still open: fact visibility is not enforced on the canonical KB search paths. Pointing the import at the real memory directory would make the owner's whole memory directory (hundreds of files) searchable by any user until that lands. Redaction covers 7 detector categories, not internal context.
Cheapest discharge that avoids both:
- Create a directory readable by the service user holding one synthetic
*.mdmemory file with frontmatter, and setAUTOBOT_CLAUDE_MEMORY_DIRto it for the backend service (a config change to the live service, so through the builtin path, not a shell edit). - As an admin, record the knowledge_facts count,
POST /import_claude_memory, poll its status, record the count again. - Search the KB for a phrase from the synthetic file and confirm the fact is returned with
category="claude_code_memory".
That discharges the one open criterion and the
blocksedge into #16661 (native edge confirmed: #16661 isblocked_by#16642).Batches with #16280: one admin session.
- Create a directory readable by the service user holding one synthetic
Context
AutoBot already has a memory architecture (
autobot-backend/memory/):EssentialStoryGenerator(compact always-loaded summary),WorkingMemoryService(Redis, session-scoped, TTL),GeneralStorage/TaskStorage, andknowledge_facts(Postgres system-of-record, Redis+ChromaDB projections, declared inautobot_shared/store_authority.py).Separately, Claude Code sessions on dev machines maintain an auto-memory store at
~/.claude/projects/<project-slug>/memory/*.md— each file has frontmatter (name,description,metadata.type: user|feedback|project|reference) plus a body. This is curated project knowledge (decisions, gotchas, process feedback) accumulated across many coding sessions, currently invisible to AutoBot's own agents/chat.Goal
A reusable, idempotent importer that reads Claude Code auto-memory files and writes them into AutoBot's existing
knowledge_factsstore (via the existingknowledge.facts/FactsMixinwrite path — no new store, no bypass of the declared system-of-record), so AutoBot's own KB/RAG can surface this knowledge.Location
autobot-backend/knowledge/facts.py(FactsMixin)autobot_shared/store_authority.py(knowledge_factsconcept)autobot-backend/api/knowledge_population.py(check_admin_permission)Acceptance Criteria
<memory-dir>/*.mdfrontmatter + body; memory directory path is configurable (SSOT/env), never hardcodedknowledge_factsentry tagged with itsmetadata.typeand a source marker (e.g.category="claude_code_memory")knowledge.factswrite path declared instore_authority.pyforknowledge_factscheck_admin_permission, consistent withknowledge_population.py)Out of Scope