memory-engineering

Run memory_density_auditor.py with --dir pointing at your memory directory or --jsonl for a records file. It classifies each record as FACT, SKILL, LOG, or PROSE, finds near-duplicates above 75% similarity, and flags records older than a configurable staleness threshold.

Run memory_architecture_picker.py with a constraints JSON describing latency sensitivity, build budget, recall need, volume growth, and mutability. It scores the four paradigm families, disqualifies any that violate hard constraints, and names the cost the winning choice makes you pay.

The Stanford characterization found construction energy exceeds total query energy across 300 queries for LLM-mediated memory systems. memory_cost_profiler.py reports this split and computes cost per correct answer so you can cut construction tokens before tuning retrieval.

No. All four scripts are stdlib-only Python with no network access, no LLM calls, and no third-party dependencies. They perform deterministic arithmetic, lexical classification, and policy linting entirely locally.

Two blocking checks fail a design: F1 requires an explicit forgetting rule (TTL, capacity bound, or decay), and F4 requires contradictions be surfaced to a human rather than auto-merged. Other gaps like missing audit trails or rollback degrade the verdict to CONDITIONAL.

Run each pass by hand against real history first and ask whether it changed a decision. If the output did not change anything, scheduling it only generates noise, and sparse input causes the system to hallucinate connections that are not there.