Engram documentation

Capture memory operations, explain bad agent answers, replay controlled repairs, and preserve the fix as a regression test.

Engram is an open-source memory reliability workspace for AI agents. It helps you answer one concrete engineering question:

This agent gave a bad answer. Which memory decision caused it, what changes if I correct that decision, and how do I prevent the failure from returning?

Capture your first turn

Install the CLI and SDK, run your agent with local capture, and inspect the resulting turn.

Understand the lifecycle

Learn the difference between storing, retrieving, loading, updating, and consolidating memory.

Diagnose an incident

Connect a bad answer to the exact memory evidence available at each stage.

Run the reference failure

Reproduce a stale-location answer, repair it on a branch, and execute the exported regression.

The core loop

1. Observe

   Capture what the memory system stored, updated, considered, selected, and loaded into model context.

2. Diagnose

   Find the earliest recorded memory decision that can explain the bad answer without inventing missing evidence.

3. Intervene

   Change memory state, eligibility, ranking, selection, or context on an isolated branch while preserving the source run.

4. Replay

   Reproduce the baseline, rerun every supported stage, and compare the first observable divergence.

5. Prove

   Export semantic lifecycle and answer assertions, then run them across paraphrases, entity changes, score ties, and distractors.

Start locally

npm install --save-dev @engramviz/cli
npm install @engramviz/sdk
npx --yes @engramviz/cli init --project my-agent
npx --yes @engramviz/cli dev

Then run your agent with the capture environment injected:

npx --yes @engramviz/cli run -- npm run my-agent

Engram Studio opens locally at http://localhost:3100/?mode=incidents.

Engram visualizes observable application behavior, not hidden chain-of-thought or model activations. A loaded memory was available to the model; availability alone does not prove that the model relied on it.

How the pieces connect

The 3D brain is a synchronized educational and evidence map. The Incident workspace is the engineering surface for diagnosis, intervention, replay, and verification.