Agent Memory Manager

Verify

VERIFY — agent-memory-manager

How verification runs and what each check asserts. The harness is verify.mjs; the unit suite is under test/ (compiled to dist/test).

How it runs

node verify.mjs performs, in order:

Build — compiles TypeScript with tsc.
Unit suite — runs node --test over the compiled tests and asserts zero

failures.

Synthetic corpus — a seeded PRNG (mulberry32, seed 20260625) builds N

topics, each with a unique vocabulary, so same-topic texts share tokens (high cosine) and cross-topic texts do not. This yields a known retrieval answer key.

Dual-store benchmark — the same corpus is loaded into the InMemoryStore

and a live PGlite Postgres+pgvector engine; retrieval precision@k is measured against the answer key on both.

Policy/behaviour checks — decay demotion, tier eviction, fleet sync,

determinism, and metrics.

What each check asserts

Check	Assertion
Unit suite passes	`node --test` reports 0 failing tests (24 tests).
In-memory precision@1 == 1.0	Top hit for each topic query belongs to that topic.
In-memory precision@5 ≥ 0.95	≥95% of the top-5 belong to the queried topic.
Deterministic retrieval	Re-running the same query yields identical ordering.
pgvector precision@1 / @5	Same thresholds, computed over real pgvector.
pgvector vs JS cosine	`1 - (embedding <=> q)` agrees with the JS cosine within 1e-5.
pgvector persistence	Row count in the DB equals the number of stored memories.
Store parity	In-memory and pgvector pick the same top-1 topic for every query.
summarizeThread	Produces a non-empty summary and persists a `summary`-typed memory that is retrievable.
getContextForTask budget	Assembled context respects the character budget and ranks the summary first.
Time decay demotion	A hot memory whose decayed importance drops below the hot floor is demoted to warm.
Hot capacity	After overflow, the hot tier holds ≤ its configured capacity.
Cold eviction	The cold tier is capped and `memory_evict_total{tier=cold}` is emitted.
Fleet sync	Fleet-scoped memories replicate to a peer; local-scoped ones do not.
Metrics	Counters and latency histograms are recorded.
Reproducible corpus	The same seed regenerates an identical corpus fingerprint.

Outputs

verification-report.json / .md — machine- and human-readable results.
evidence/verification-results.json — copy of the report.
proof/evidence/verify.log — raw stdout captured by the Proof Layer.

A non-zero exit code is returned if any check fails.