[idapixl]

Why

#147. Abstract handed the model one memory from each of 3–4 random categories and asked what connects them. The model always found something by climbing to a theme ("friction and identity", "subtraction over addition", "agency and constraint"). One live run linked a failed MCP integration to a person's family history, and each result was stored at salience 0.8, above most organic memories.

What

• Related members. Each attempt seeds on a recent memory. Its members are the seed's graph neighbours (any relation except exemplifies) plus near neighbours at cosine ≥ abstraction_relate_floor (default 0.50), preferring categories the group doesn't have yet. When fewer than two related members exist, the model isn't asked. Existing abstractions are never members.
  • Measured on one store (qwen3-embedding): random pairs p90 0.43, edge-linked pairs median 0.37–0.42. The graph carries the signal; 0.50 admits only near neighbours.
• A concrete shared mechanism. abstract-synthesis v3 asks for JSON (mechanism, members, name, abstraction), with "none" as the expected answer and a rule against linking a person's private life to a technical system through a theme. The new checkMechanism (engines/abstraction-check.ts) requires the mechanism to contain a term that at least two distinct cited members' own text uses:
  • not an abstract noun;
  • in pools of 20 or more, in no more than 25% of memories;
  • members that restate one fact (term Jaccard ≥ 0.25; edge pairs p99 0.294, the Muzak restatement pair 0.308) count once.
• Salience 0.5 (ABSTRACTION_SALIENCE), down from 0.8. exemplifies edges go only to cited members, and each edge's evidence names the shared terms. phases.abstract.rejected is reported.

Measurement (copies of one 694-memory store, packed build of this branch)

• Copy dream, first build: 1 abstraction written, a restatement of two Muzak memories. That led to the distinct-members rule.
• Copy dream, final build: 0 damaged, 0 regressions; abstract wrote 0 and rejected 1.
• 20-attempt probe of the phase alone: 18 "none", 1 restatement rejected, 0 written.
• Same probe without the "most groups have none" priming line: 8 written, 9 rejected. About half of the 8 were theme-glue under concrete-sounding names, including one tying code optimisation to a memory about money.

I kept the conservative prompt: a phase that will run unattended should be quiet, not wrong. See the CHANGELOG entry.

471/471 tests. 4 of the new phase tests fail on master, and the rest pass there only because master never calls generateJSON.

Closes #147

🤖 Generated with Claude Code