KnowledgeStore-Guided Hippocampus Prefix Extraction — 2026-06-27

Question

Can the hippocampus-derived ReservedPosition path benefit from the KnowledgeStore when a recalled fact does not have a hand-seeded reserved_token?

Method

Added KnowledgeStore.get_prefix_targets(query) which returns target strings from the top recalled facts. Preference order:

1. metadata["prefix_target"] if present.
2. metadata["reserved_token"] if present.
3. The fact "value" otherwise.

Facts must meet the same min_score threshold used by format_context() and get_reserved_position().

Added CortexAgentConfig.knowledge_store_supplies_prefix_targets (default False). When enabled, CortexAgent.articulate() collects targets from the KnowledgeStore and passes them as prefix_targets to _derive_reserved_position_from_hippocampus() whenever no explicit ReservedPosition was already set by get_reserved_position().

Added tests:

• test_knowledge_store_get_prefix_targets: verifies values and explicit prefix_target metadata are returned.
• test_hippocampus_prefix_uses_knowledge_store_targets: mock driver shows a non-reserved-token fact still yields a ReservedPosition via the hippocampus helper when the flag is enabled.

Results

• KnowledgeStore unit tests: 16 passed.
• CortexAgent tests: 15 passed.
• CortexAgent reserved-position tests: 4 passed.
• Fast suite: 315 passed, 26 deselected.
• ruff check clean.
• Benchmark code_qa_accuracy=1.0 (run #102).

Interpretation

The KnowledgeStore is no longer limited to exact-token steering only for facts with reserved_token. Any recalled fact can now guide hippocampal snippet-extraction via prefix_targets, extending exact-token recall potential to the broader codebase-QA corpus.

The flag is default-off, so existing behavior is unchanged. When enabled, the precedence remains: explicit ReservedPosition > hippocampus-derived prefix > none.

Limitations

• The benchmark was already at code_qa_accuracy=1.0; no improvement was observed, because the existing reserved-token facts cover the questions.
• A real workload showing improved recall on facts without reserved_token has not been run.
• get_prefix_targets currently returns the top fact(s') values; a long fact value could produce a very long prefix if the truncation window also captures filler text.

Next steps

1. Run codebase-QA with knowledge_store_supplies_prefix_targets=True and compare recall_lift.
2. Measure IdentityHypernetwork adapter effects on the multi-fact probe.
3. Close the benchmark gap on exact-token consolidation uptake.

Artifacts

• src/oczy/experiments/codebase_qa/knowledge_store.py
• src/oczy/experiments/cortex_agent.py
• src/oczy/experiments/codebase_qa/test_knowledge_store.py
• src/oczy/experiments/tests/test_cortex_agent.py

Commits

• 311cf19 — Integrate KnowledgeStore prefix_targets.
• b1311d6 — Update SUMMARY.md with run #102 result.

Run

Run #102: benchmark code_qa_accuracy=1.0, fast suite 315 passed.