Replay and regression-test agent memory policies.

Import a captured failure, inspect the exact state, ranking, selection, and context decisions, then test an alternative policy on an isolated branch and keep the repair executable in CI.

Run the same incident locallynpm

npx --yes @engramviz/cli demo stale-location

Open-source CLI Local Studio No account No hosted collector

Your agent → SDK or adapter → .engram/data → local Studio

Run the guided incident Instrument your agent

From trace evidence to a memory-policy test.

Diagnose one stale-memory incident, replay a controlled repair, then make it portable.

Observed

Find the first bad memory decision.

Engram follows the recorded decision from memory state to answer.

Frozen turnWhat city do I live in now?

Stale stateSan Franciscostill active

SelectedSan Franciscoloaded

Use your existing memory stack.

Capture retrieval, loaded context, and answers with the SDK or an adapter.

TypeScript SDK@engramviz/sdk Mem0@engramviz/adapter-mem0 LangGraph Store@engramviz/adapter-langgraph Custom providerMemory Telemetry v2

Evidence without invented certainty.

Engram shows observable application behavior, not chain-of-thought or hidden model activations. Replay is behavioral evidence, not causal proof. Missing evidence stays unavailable instead of becoming a confident story.

Use Engram beside general tracing tools as the specialized memory-policy layer: reconstruct the lifecycle decision, intervene on it, and turn the repair into a semantic regression contract.

ObservedDerivedReplayedUnavailable

Instrument your agent