Remember. Measure. Reconsider.

A place for experience to persist.

UniMem is evidence-based memory infrastructure for AI agents. It stores reusable lessons, time-aware facts, and other forms of context, and connects memory use to task outcomes.

Lessons capture reusable guidance. Atomic facts carry both valid time and record time, so a workflow can distinguish what was true from what the store believed at a particular point. Conversations, memory blocks, and knowledge graphs serve different kinds of continuity.

The local core uses JSONL, Markdown, and SQLite. Start without an embedding service or an LLM; optional providers add semantic retrieval and reranking. A vector index helps find information, while admission and lifecycle controls determine whether it may be used.

Usefulness is a testable question.

A memory appearing in successful tasks is not enough to show it helped. It may simply appear in easy tasks—or be retrieved most often when the task is unusually difficult.

In a configured experiment, UniMem assigns eligible memories to shown or withheld arms for each task. A stable occasion ID links assignment to the success or failure reported by your workflow. Repeated retrieval on the same occasion does not become a new independent task.

The report compares outcomes and checks experiment integrity. Results distinguish evidence of help, evidence of harm, no measurable effect, and an underpowered experiment. The last of these means the experiment cannot answer yet.

• Define the task outcome before interpreting the trial.
• Honor withheld memories and report outcomes for every assigned occasion.
• Keep memory revisions, experimental design, and outcome scoring consistent.
• Reserve protected safety guidance from ordinary memory experiments.

Evidence can inform an explicit change.

With a configured harm policy, UniMem can stop delivering memories with measured evidence of harm. That is different from assuming every unhelpful-looking memory should be deleted.

A separate change-control workflow preregisters candidate and incumbent policies, support requirements, and an acceptance margin before observing outcomes. Certification requires canonical trial records and a passing integrity audit. Rollback records revoke the certificate.

Lesson admission also has a review path. Content-bound receipts connect approval to the reviewed revision. Two-person and human-required review are deployment choices your operator must configure.

Fit the loop around your workflow.

Use the CLI for a first local store, MCP for an attached agent, the HTTP gateway for a language-independent interface, or Python to wrap a retrieval callable. The measurement wrapper can work around an existing store when returned memories have stable identities.

A useful first integration has three responsibilities: retrieve admitted candidates, give the agent only the delivered set, and record the final outcome. UniMem cannot manufacture a trustworthy success signal for your business process.

The contextual causal gate is newer functionality. The source documents that the Hub does not yet retain all per-assignment rates and contexts required for the local gate. Evaluate this boundary before choosing a deployment.