AI/Tasks/MemoryUpgrades/Memory_Phase07_LearningAccuracy_Strict.txt
|
STRICT MODE: Improve learning accuracy in the memory subsystem.
Mission: - Complete this phase with bounded steps and deterministic fallback behavior. - Apply changes ONLY within the memory subsystem and its tests. Concrete scope: - src/TechToolbox.Agent/Memory/MemoryLearner.cs - src/TechToolbox.Agent/Memory/MemoryStore.cs - src/TechToolbox.Agent/Memory/MemoryModels.cs - src/TechToolbox.Agent/Tests/MemoryLearnerTests.cs (plus additional memory tests if needed) Requirements: - Distinguish durable user preferences from one-time task instructions. - Replace broad regex-only inference with structured evidence metadata where possible. - Require stronger evidence thresholds for permanent factual memory entries. - Persist source category and confidence reason/rationale for learned items. - Add adversarial tests: quoted text, negation, hypotheticals, examples/counterexamples. - Do not modify unrelated subsystems. Step budget (hard limits): 1) Discovery: <= 6 read/search calls. 2) Edits: <= 9 file edits total. 3) Validation: 1 build + 1 targeted test run; one repair cycle max. 4) Stop once criteria pass. Loop guards: - No third pass on heuristic wording tweaks without new evidence from tests. - If adversarial tests keep failing after one repair cycle, stop and report pattern-level blocker. Fallback behavior: - Fallback A: implement conservative durable-memory gate first (high precision, lower recall). - Fallback B: if structured metadata rollout is too broad, store minimal provenance fields now and document extension path. - Fallback C: after one failed repair cycle, emit FAILED_VALIDATION. Acceptance criteria: - One-off instructions are not promoted to durable preferences by default. - Learned-item provenance and confidence reason are queryable. - Adversarial parsing tests pass and prevent common false positives. Validation commands: - dotnet build src/TechToolbox.Agent/TechToolbox.Agent.csproj -c Release - dotnet test src/TechToolbox.Agent/Tests/TechToolbox.Agent.Tests.csproj -c Release --filter "MemoryLearner|Memory|Preference" Output results in markdown using this structure: 1) Summary 2) Files changed 3) Learning classification rules 4) Evidence/provenance model 5) Tests added/updated 6) Validation results 7) Step-budget report (planned vs actual counts) |