“The service listens on port 3000.”
Observed July 2The problem was not storage.
An agent can keep writing notes forever and still have bad memory. The harder questions are whether a request needs memory, which evidence deserves context, whether it was useful after retrieval, and what should happen when it becomes stale or wrong.
Cortex turns those questions into an inspectable local system. The core works beside any agent harness, keeps its database on the agent’s host, and adds no extra model call to the normal recall path. Hermes remains the first reference adapter.
Health without graph literacy.
A memory graph is useful for exploration, but it should not be required for maintenance. The Brain dashboard turns review work into one plain-language decision at a time.
GUIDED REVIEW · 1 OF 12Which statement should Cortex use now?
Safe to use“The service listens on port 3001.”
Observed July 12A is currentB is currentBoth are validNot sure
Every choice has a confirmation step. Cortex versions or archives the old belief, preserves provenance, and records an audit trail.
A bounded recall loop.
The hot path is deterministic so it can be measured, debugged, and replaced piece by piece.
01
Gate attention
Skip memory on social and self-contained turns; use deeper plans for durable, temporal, and multi-memory work.
02
Find evidence
Combine FTS5, transparent semantic features, and a bounded walk through explicit associations.
03
Budget context
Rank by relevance, time, source trust, utility, harm, and uniqueness; then stop at a small prompt budget.
04
Credit actual use
Retrieval alone earns almost nothing. Structured values, distinctive anchors, and later outcomes determine utility.
05
Repair over time
Version corrections, learn repeated tool workflows, consolidate duplicates reversibly, and track pruning regret.
Psychology as a constraint, not a costume.
Bounded access
Working-memory research motivated small, task-sensitive recall—not a claim that tokens are human chunks.
Activation
Recency and meaningful reuse affect accessibility, while harm and false positives push the other way.
Association
Cues can bring related evidence into reach through explicit, capped graph activation.
Reconsolidation
Corrections create versions and changed evidence makes dependent beliefs reviewable.
Consolidation
Episodes stay raw while repeated evidence can become a stable fact or tool procedure.
Adaptive forgetting
Low-value memories cool and archive to reduce interference, with restoration when pruning regret appears.
What the evidence says today
Much better recall in the first paired test. No demonstrated TTFT win—yet.
Across 90 paired synthetic questions against 500 memories, the additive Cortex condition answered 96.7% correctly versus 6.7% for Hermes’s bounded built-in snapshot. Time to first token was effectively tied. The original fixed-recall Cortex run used more prompt tokens; the 0.2 local ablation now cuts approximate memory-context tokens by 21.7% without reducing labeled answer availability in that sample.
Synthetic benchmarks are not a universal result. The repository includes raw aggregates, uncertainty, scripts, and explicit rules for what can and cannot be claimed.