Files
dotfiles/claude/memory/bmf-summarizer-contract.md
T
David F GliddenandClaude Opus 4.8 3f9a89b00c chore(memory): Basic Memory trial begins — sync normalization baseline (283 files)
Basic Memory v0.21.6 first sync over the live memory dir (steward-authorized
live-dir trial, Option A 2026-06-06): adds permalink: to frontmatter, refolds
long YAML description lines, strips final newlines. Bodies untouched —
verified via full diff classification. From this commit forward, any diff in
claude/memory shows only what Basic Memory or the session writes.

Trial design: MemPalace untouched as incumbent; git status check on this dir
at every wrap; end-of-day evaluation (recall quality, sync robustness,
rebuild-from-files, malformed-file behavior).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-06 09:52:17 +02:00

3.4 KiB

name, description, type, originSessionId, permalink
name description type originSessionId permalink
BMF summarizer-connector coupling Factory's maybeGenerateSummary is silently coupled to the local_files observation shape; most connectors fall through. Deferred refactor target. project 1c1bf37d-b1c3-4a62-9355-613c50f4c028 claude-memory/bmf-summarizer-contract

BMF summarizer-connector coupling (deferred refactor)

Surfaced: 2026-04-14, during the SQLite+LanceDB recall unblock session.

src/factory/factory.ts::maybeGenerateSummary implicitly requires the local_files observation shape:

  • source.item_type ∈ {'file','image','passage'}
  • payload.metadata.file_path (or source.item_id)
  • payload.metadata.mime_type (unless item_type is 'file' or 'image')
  • payload.metadata.size_bytes
  • source.content_hash

When documentSummarizer.isAvailable() is true, processFactoryBatch previously set skipDispatch: true uniformly for the whole batch. Any observation that didn't match the summarizer's entry guard was written to logchain and then silently dropped — never dispatched, never summarized. This was the actual cause of zero embeddings after 2,457 Obsidian notes ingested.

Audit of mime_type usage across connectors (2026-04-14):

  • Sets it: local_files, google, obsidian (post-fix).
  • Does NOT: claude_transcripts, slack, github, apple_calendar, apple_contacts, imessage, git, file_import.

Every connector in the second group drops content silently when a summarizer is available — until the partition fix landed.

Why: fixes applied in this session

  • Structural safety net (factory.ts::processFactoryBatch): partition the batch by a summary-eligibility predicate mirroring maybeGenerateSummary's entry guard. Summarizable → skipDispatch: true; pass-through → skipDispatch: false. Guarantees every content type reaches modules regardless of summarizer shape.
  • Notes first-class (connectors/obsidian.ts, factory.ts::maybeGenerateSummary): emit mime_type: 'text/markdown' + file_path; extend summarizer 'file' branch to accept 'note'. Notes get the quality path (summary, teacher enrichment eligible).
  • Provenance (factory.ts::generateAndObserveSummary): connector_id and factory:* tag now use job.connector_id instead of hardcoded 'local_files'.

How to apply: the remaining [PROPOSAL]

Not "add mime_type everywhere." Most connectors (git, apple_contacts, apple_calendar, slack/imessage messages, claude_transcripts per-turn, short github comments) are correctly pass-through and should NOT be summarized. The right generalization is an explicit contract:

  • Add summarizable: boolean to ConnectorObservation (or equivalent) — chosen per-connector.
  • Dispatch partition uses the flag directly rather than inferring via item_type/mime_type.
  • Summarizer accepts a generic-text shape (title + content + content_hash), not local_files-shaped metadata.
  • Fail-loud telemetry: a batch with zero summaries AND zero pass-through dispatches should alarm, not silently succeed.

Touches connector contracts — deserves its own focused PR with Seb reviewing, not a tail-end addition to the storage-migration PR.

When to revisit

  • When opening the next L1 PR after the storage-migration one merges.
  • Before onboarding any new text-bearing connector (the wrong default will keep catching people).
  • If any recall quality investigation points at content-not-dispatched as a suspect — this is the mechanism.