Memory Proposal Auto-Triage Pipeline

Context

The memory reflection pipeline (ADR-0021) generates proposals from session observations and queues them for human review. In practice, this creates an unbounded backlog (88+ pending) that never gets reviewed because:

• Most proposals are factual observations that don’t need human judgment
• Many are duplicates across sessions (same insight observed 3+ times)
• Instruction-text artifacts leak into proposals (“Add after…”, “Replace…”)
• No priority signal — trivial version bumps sit next to critical debugging insights
• Review is one-at-a-time via joelclaw review — tedious, low-value

The result: proposals accumulate, nothing gets promoted, MEMORY.md stagnates, and the system learns nothing from its own observations.

Decision

Replace the flat review queue with a three-tier auto-triage pipeline:

Tier 1: Auto-Action (no human needed)

Auto-promote when ALL of:

• Contains a date reference (factual, timestamped)
• References a specific tool, file, command, or config
• Does not conflict with an existing MEMORY.md entry (embedding similarity < 0.85)
• Is not a preference/opinion statement

Examples: tool installed, config changed, bug found + fix documented, new service deployed.

Auto-reject when ANY of:

• 85% embedding similarity to existing MEMORY.md entry (duplicate)

• 85% similarity to another pending proposal (dedup)

• Content starts with instruction text (“Add after”, “Replace”, “Expand”, “Consolidate”)
• References a superseded ADR or removed tool
• Is a raw - (YYYY-MM-DD) line (leaked format)

Auto-merge when:

• 2+ proposals target the same MEMORY.md section with overlapping content
• Keep the most comprehensive version, discard others

Tier 2: Batched Human Review (weekly)

Proposals that survive Tier 1 but aren’t auto-promotable:

• Preference changes (“Joel wants X instead of Y”)
• Architectural opinions
• New conventions or hard rules
• Anything ambiguous

Presented as a batch (5-10 items) via gateway digest:

• Grouped by MEMORY.md section
• Each has a recommendation (promote/reject) with reasoning
• Joel confirms or overrides per-item
• 2-minute weekly task, not 20-minute daily grind

Tier 3: MEMORY.md Hygiene (monthly)

• Flag entries referencing old tool versions, completed projects, or superseded ADRs
• Propose removals/updates as a batch
• Prevents MEMORY.md from growing unbounded

Weekly Digest

Gateway message every Monday:

Memory Triage: 12 auto-promoted, 8 auto-rejected, 3 merged, 5 need your call:
1. [Conventions] "Repos should use X pattern" — recommend: promote
2. [Hard Rules] "Never do Y in Z context" — recommend: promote
3. [Joel] "Prefers A over B" — recommend: review (new preference)
4. [Patterns] "Z pattern works for W" — recommend: reject (already covered)
5. [System Architecture] "Service X moved to Y" — recommend: promote

Implementation

1. New Inngest function memory/proposal-triage triggered by memory/proposal.created
2. Runs embedding similarity against MEMORY.md entries (Qdrant memory_observations)
3. Classifies into auto-promote / auto-reject / auto-merge / needs-review
4. Auto-actions execute immediately with audit log
5. Needs-review proposals accumulate for weekly digest
6. joelclaw review updated to show batched, grouped proposals with recommendations
7. Weekly cron memory/digest.triage sends the summary via gateway

Backlog Cleanup

Run the 88 pending proposals through the same triage on deploy. Expected outcome: ~50 auto-rejected (duplicates + instruction text), ~25 auto-promoted (factual observations), ~13 needing actual review.

Consequences

• MEMORY.md stays current without human bottleneck
• Proposals stop accumulating — the queue drains itself
• Joel only sees genuinely ambiguous items (5-10/week)
• Embedding similarity requires Qdrant memory_observations to have indexed MEMORY.md entries (Phase 5 dependency, can use text similarity as fallback)
• Risk: auto-promote writes something wrong. Mitigation: git history on ~/Vault, weekly digest shows what was auto-promoted for spot-checking.