Memory

Memory lets an agent recover useful context after a session ends. Layr8’s memory service, mnemo, stores conversation history, facts, and saved notes so authorized agents can retrieve them later.

For a hands-on example, save a preference and retrieve it in a fresh session.

From session context to lasting memory

An agent’s current conversation provides context for the work in front of it. A new session needs a way to recover relevant decisions, preferences, and unfinished work. Mnemo provides storage and retrieval outside that conversation.

For example, an agent can save why a project chose a particular approach. A later session can retrieve that decision before proposing changes. Another authorized agent can use shared notes without needing the original conversation copied into its prompt.

Saving and retrieving are separate steps. What happens automatically depends on the launcher and its memory integration. Loading the memory skill gives an agent instructions for using mnemo; it does not by itself enable a disabled integration or ensure every conversation is saved.

Partitions organize memory

A partition is a collection of memories with its own membership. Use partitions to separate context for different projects or teams. A partition is distinct from the Layr8 Space that contains your agents and services. Some protocol and integration settings call a partition a “memory space.”

The onboarding flow uses a default partition. Follow the cross-session tutorial to check your launcher’s configuration and verify that a fresh session can retrieve what you saved.

What memory contains

Kind	Purpose	Example
Conversation history	Recover context from recorded exchanges.	The discussion that led to a decision.
Facts	Represent information that can change over time.	A project’s current deployment target.
Learnings	Keep named notes that can be updated or deleted.	A preferred progress-reporting format.

Updating a fact can supersede an earlier value while retaining its history. Marking a fact as forgotten makes it no longer current; it is not a promise that all historical records have been erased. A learning is a named value: saving under the same key updates that note.

Retrieve what matters

Agents can request saved learnings, query facts, search recorded conversations, or ask mnemo to assemble context within an estimated token budget. Assembled context can combine recent turns, older summaries, learnings, and facts.

A retrieved result is a selection, not a complete transcript of everything stored. Budgets, search relevance, permissions, and service configuration affect what comes back. Semantic search also depends on the deployment’s embedding backend.

For important decisions, ask the agent to retrieve the specific saved information and check it against the current task. Verify important writes with confirmation and readback rather than assuming a sent request succeeded.

Sharing and access

Partition membership controls access. Readers can retrieve memory, writers can also save it, and owners manage membership. Treat a partition as a shared access boundary. Some context requests filter which conversations are selected, but those filters do not provide private storage within a shared partition. Use separate partitions when information needs different membership.

Treat saved memory as context. A stored preference or decision can inform work, but it does not grant permission to act or override your current instructions. Correct outdated information when you find it, and use separate partitions where projects or teams need separate access.

Try it

Follow Use mnemo across sessions to save a preference, verify it, and recover it in a new session. Once that works, use memory for decisions, corrections, and unresolved questions that will help the next session continue.