Audit recalled memory for stale facts, contradictions, secrets and unsupported claims before assembling the context an agent will use.
A 23-second tour of the actual console follows the sample taskflow-api project through its memory graph, audited recall, reusable rules and forgetting controls.
Recorded sample project · on-screen explanations · no audioOpen recordingPersistent memory can confidently bring back the wrong thing: a superseded deployment target, an unsupported fix, contradictory advice, or a credential captured in a transcript.
I built a review layer around Cognee’s memory lifecycle. It compares raw recall with an audited context pack and gives both the person and the agent an explicit way to retract a memory.
The engineering behind the screens, in three stages.
Typed memories retain their session and evidence relationships. Secret detection runs at ingest and again on reads; the checks are bounded rather than a universal security guarantee.
The audit checks validity, contradiction, sensitivity and supporting evidence. Supersession relationships help distinguish a remembered fact from the current fact.
The lifecycle includes improving reusable rules and forgetting a selected memory. MCP clients can request audited context and operate the same memory lifecycle.
Typed nodes and supersession relationships support temporal and contradictory-memory checks beyond similarity search.
Raw recall beside the reviewed context pack explains what was removed and why. The filtering decision can be inspected.
Memory governance includes retraction, not only adding more facts to an ever-growing context store.
Architecture, recorded Cognee runs and a public build writeup.
This is a prototype with bounded checks. A passed audit does not prove that every memory is true or safe. Local storage and local inference are separate settings; the default model configuration may call a hosted provider.
An agent’s memory is a changing source of evidence, not an unquestionable instruction.