Research generic-fact supersession against external close/open fixtures

[hjqcan]

Evidence

The six public HeartFlow fixtures at yun520-1/mark-heartflow-skill@f3cc0874 provide concrete cases for same-slot replacement, aliasing, multi-valued predicates, negation, ambiguous correction, and explicit clearing.

A cross-check against the published npm package goodmemory@0.7.2 found:

• bare fixture statements: 0/12 admitted by the selective write gate, so this is not a valid supersession comparison
• explicit Remember that ... normalization: 12/12 admitted
• ordinary generic-fact remember: both facts remained active in all six fixtures
• exact fixture-level lifecycle match: 1/6 (the multi-valued coexistence case)

Reproduction, method, and limitations:
https://gist.github.com/hjqcan/012e62e4bc42e325eb69db3ac3b19eb5

Original ecosystem discussion:
MoonshotAI/kimi-cli#1283 (comment)

Current boundary

This confirms the existing documented limitation: generic explicit facts are not universally assigned to a semantic slot and atomically closed by ordinary remember. GoodMemory does have lifecycle state, targeted revision, and lane-specific supersession, but those paths are outside this cross-check.

The HeartFlow-specific metadata.type and tags were not mapped because GoodMemory’s public remember input does not accept them. This result must not be presented as an apples-to-apples product benchmark or as evidence that HeartFlow itself passes the fixtures.

Research question

Can generic close/open semantics be strengthened without false-closing aliases, additive/multi-valued facts, or ambiguous corrections?

Before changing defaults, any proposal should:

• define the subject/predicate identity available at the write boundary
• include explicit false-close and missed-close cases for alias, multi-value, negation, and ambiguous correction
• preserve the conservative behavior when referent identity is not established
• distinguish ordinary remember from explicit targeted revision
• show held-out evidence rather than tuning only to these six public fixtures

[hjqcan]

Follow-up against the contract-aligned tagless variants in yun520-1/mark-heartflow-skill@a9fa58d3, using the published goodmemory@0.7.2 package (npm integrity sha512-rDPHNJnhSqbPsSMqLm0Ri0ggif4aeHKjc6uDRPvr4M4pPviJiJi175LAf2LybffWv3wsQBgSdR33T42JK+/Xwg==) with Bun 1.3.14.

Method: verbatim natural-language content, one session per entry, a fresh in-memory workspace per fixture, and no metadata.type or tags.

Results:

• fixture-2-tagless-alias-gap.json: both “Alice Chen” and “A. Chen” remain active; no destructive supersession. Active-set and intended behavior match.
• fixture-5-tagless-ambiguous-correction.json: both “blue” and “red” remain active, so there is no false overwrite; however, GoodMemory emits no conflict flag, proposal, or review signal. The active set matches, but the full intended behavior does not.

Summary: active-set match 2/2; full behavioral match 1/2.

Compact raw result

{
  "package": "goodmemory@0.7.2",
  "fixtureCommit": "a9fa58d3",
  "method": "verbatim content; separate session per entry; fresh in-memory workspace; no metadata.type/tags",
  "results": [
    {
      "fixture": "fixture-2-tagless-alias-gap.json",
      "writes": ["accepted", "accepted"],
      "active": ["mem-t2-001", "mem-t2-002"],
      "proposalCount": 0,
      "fullMatch": true
    },
    {
      "fixture": "fixture-5-tagless-ambiguous-correction.json",
      "writes": ["accepted", "accepted"],
      "active": ["mem-t5-001", "mem-t5-002"],
      "proposalCount": 0,
      "fullMatch": false,
      "gap": "no conflict/review signal"
    }
  ]
}

Boundary: this is a focused contract comparison, not a product benchmark, coding-outcome claim, or adoption signal.