Applied use case / Memoato

Memory that keeps the evidence in view.

Memoato is the open-source product behind our research into trustworthy AI memory and context. It is currently a personal memory product: a place to record what happened, review what the system understood, and recall it with the source still attached.

Product
Personal memory
Canonical store
PostgreSQL
Agent access
Logging, recall, or both
Status
Live product, active development
Memoato landing page showing a saved personal note and the details remembered from it
Memoato starts with ordinary language and keeps the original entry available beside the details derived from it.

Three connected layers

Research, product, and implementation work inform each other.

01

Trustworthy Agent Memory

The public guide develops principles for provenance, freshness, permissions, correction, and evaluation in agent context systems.

02

Memoato

The open-source product puts those principles under real product constraints: capture, processing, review, retrieval, and external access.

03

HILLS Lab

We apply the same engineering discipline when helping software teams design and implement context-aware products and workflows.

The trust problem

A generated memory is not automatically a fact.

A short personal note can contain an event, a reason, a preference, and an uncertain interpretation at the same time. Flattening all of that into one generated summary hides what the person actually wrote and makes later correction difficult.

Memoato keeps the source entry before interpretation. Structured facts are reviewable records, while inferences stay separate and point back to evidence. The system can improve its interpretation without rewriting the original account.

Record model

Each kind of knowledge has a different job.

Raw evidence

The original entry is preserved as the durable source. Later processing can be rerun without replacing the user's words.

Reviewed facts

Structured facts have their own review state. External recall is limited to accepted facts rather than every generated candidate.

Separate inferences

Patterns and hypotheses remain distinct from facts and carry references to the entries that support them.

Corrections and runs

Human corrections and processing attempts are recorded, leaving a visible path from source to current interpretation.

Scoped API and MCP access

An integration receives only the capability it needs.

Memoato keys are created for logging, recall, or both. The same boundary applies whether an agent connects through the HTTP API or MCP.

01

Logging

Add a raw memory entry without receiving access to recalled personal facts.

02

Recall

Retrieve accepted facts and their evidence without gaining a write capability.

03

Both

Use capture and recall together only when the integration genuinely needs both paths.

Evidence-first retrieval

A result should show why it was returned.

Recall combines structured facts with search over the preserved entries. The response keeps evidence available, so a person or agent can inspect the source instead of treating a compact answer as self-authenticating.

Search embeddings are versioned projections, not the source of truth. They can be rebuilt from PostgreSQL while the raw entries and reviewed facts remain canonical.

Product boundary

What is live and what comes next.

01

Live now

Personal memory is the first workspace. Capture, processing, fact review, corrections, separate inferences, evidence-backed recall, rebuildable search, and scoped API/MCP keys are implemented foundations of the product.

02

Direction, not a shipped claim

Team and project context will require workspace roles, source-level permissions, immutable source versions, claim precedence, freshness rules, and retrieval traces. Memoato's current model provides a base for that work; it is not presented as a finished enterprise context platform.

Explore the work

Product, source, research, and engineering note.

HILLS Lab

Designing memory or context for a software product?

We can map the sources, trust boundaries, review model, and access rules, then implement one bounded workflow with your team.