Home / Blog / Governed agent memory

What is governed agent memory?

Memory starts as a product feature. Once an agent can act, memory becomes part of the control plane.

2026-08-08 · Lians team · 8 min read

Definition

Useful recall is only the beginning.

Governed agent memory is a durable context layer that keeps facts current while preserving their provenance, validity, access boundary, and decision impact. It can answer what is true now, what was knowable at a prior moment, who was permitted to see it, and which actions relied on it.

That definition matters because a production agent does more than continue a conversation. It retrieves records, selects tools, applies policy, recommends outcomes, and sometimes triggers side effects. The moment stored context can influence one of those actions, memory is no longer only a personalization system. It is evidence.

The control test

Six questions expose the difference.

1. What entered memory?

The source, version, event time, ingestion time, subject, and admission decision should survive beyond the prompt that created the fact.

2. What was current when the agent acted?

Present recall should exclude superseded facts. Historical recall should exclude facts and revisions learned later.

3. Who was allowed to see it?

Tenant scope, information barriers, workload identity, and permissions belong in the retrieval boundary rather than in an informal prompt instruction.

4. What evidence shaped the action?

The model, instructions, recalled facts, tool results, policy version, approvals, and declared capture gaps should be bound to one portable decision receipt.

5. Was the action permitted?

A trusted receipt, authenticated identity, required scope, review status, and exact downstream request should be checked before a consequential side effect receives a permit.

6. What must be reviewed when something changes?

A policy correction or source revision should reveal the directly affected and reachable decisions, their owners, and the evidence needed to close remediation.

The Lians control loop

Memory, evidence, action, investigation.

Lians closes these questions on one append-only record layer. Each part has a separate job:

Where Lians fits

It complements memory stores, graphs, and observability.

A personalization memory can remember a user's preferences. A temporal context graph can model how facts and relationships change. An observability platform can trace the calls an application made. Lians focuses on the evidence boundary across those systems: what the agent was shown, what it was allowed to use, what action followed, and how that decision can be reconstructed later.

You can adopt that boundary without replacing the rest of the agent stack. Lians accepts native API and SDK calls, MCP tools, and OpenTelemetry evidence. It supports local, self-hosted, private, and managed deployment paths.

Available now

The category is backed by shipped software.

$ pip install lians-sdk==0.5.0
$ npm install @lians-ai/lians@0.5.0
$ uvx --from 'lians-sdk[mcp]' lians-mcp

Version 0.5.0 is available from PyPI and npm. The public Apache 2.0 repository includes the memory layer, Decision Receipt schema and verifier, Recorder, Gate, Investigator, SDKs, deployment assets, evidence mappings, tests, and reproducible benchmark materials.

Inspect the source ↗ · Open the docs → · Run the integration proof →

Claim boundaries

Governed does not mean automatically compliant.

Lians provides technical controls and evidence. It does not certify an organization, make a deployment compliant by installation, prove unrecorded context, or turn a trace into causal certainty.

Published quality and token results apply to the dated workloads, methods, and gates described with them. They are not universal savings guarantees. Private networking, KMS configuration, retention, role design, approvals, monitoring, and legal obligations remain deployment and organizational responsibilities.

Start with the boundary you have

One workflow is enough.

Developers can start locally with the SDK or MCP server. Platform teams can run the synthetic integration proof. Teams evaluating a consequential workflow can use the fixed-scope AI Evidence Readiness Sprint to map one evidence boundary and produce a clear stop, expand, or implement decision.

Start with open source → · Test a synthetic sample → · Review the paid sprint →

Give agent memory a control boundary.

Start building →