Skip to content

[adapter] Optional Lians backend for durable point-in-time forensics #53

Description

@Lians-ai

Framework or backend

Lians open-source bitemporal memory and record layer

How you'd use it

Agent Memory Guard already provides policy enforcement, snapshots, and rollback. An optional Lians-backed store could preserve guarded memory changes outside the application process so investigators can reconstruct what the guard could read at a specific decision time, retain evidence after rollback, and distinguish when a fact was valid from when it was recorded.

The target workload is a regulated or security-sensitive agent that needs durable evidence across process restarts. Network use would remain opt-in. The existing InMemoryStore would remain the default.

Security value:

  • Durable evidence survives process restarts
  • Point-in-time reconstruction separates past knowledge from later corrections
  • Deletion can be represented as a tombstone so earlier state remains available for forensics
  • Rollback does not need to erase later evidence
  • Existing policy decisions and SecurityEvent records can link to the corresponding persisted memory version

I am a co-founder of Lians and would be happy to draft the adapter, tests, and documentation if the direction fits the project.

Target API surface

Implement the existing synchronous MemoryStore contract:

  • get
  • set
  • delete
  • keys
  • items
  • contains

A first version could use one Lians agent or namespace per guarded store, persist the AMG key and serialized value with security metadata, and represent deletion with a tombstone. Unit tests can use a fake client, with one optional integration test against local Lians.

Historical reconstruction could remain an adapter-specific helper because it is outside the current MemoryStore protocol. The integration should be optional through an extra and should not change core package behavior.

Can you help?

  • I can contribute the implementation
  • I can review a PR for this adapter
  • I can test against a real workload

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestintegrationFramework integration (LangChain, LlamaIndex, CrewAI)

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions