ReMe¶

agentscope-extensions-reme integrates with the self-hosted ReMe memory service. Its distinguishing features are trajectory-based memory extraction and workspace-level isolation.

When to use¶

You want a lightweight self-hosted memory service that’s easy to spin up.
You care about summarizing whole conversation trajectories rather than individual messages.
You can express logical workspaces by userId (one workspace per user).

Add the dependency¶

<dependency>
    <groupId>io.agentscope</groupId>
    <artifactId>agentscope-extensions-reme</artifactId>
    <version>${agentscope.version}</version>
</dependency>

Quickstart¶

import io.agentscope.core.memory.reme.ReMeLongTermMemory;

ReMeLongTermMemory memory = ReMeLongTermMemory.builder()
    .userId("task_workspace")            // Maps to ReMe's workspace_id
    .apiBaseUrl("http://localhost:8002") // Your ReMe server
    .build();

ReActAgent agent = ReActAgent.builder()
    .name("Assistant")
    .model(model)
    .longTermMemory(memory)
    .longTermMemoryMode(LongTermMemoryMode.BOTH)
    .build();

userId is mapped to ReMe’s workspace_id — the smallest unit of memory partitioning in ReMe.

How it works¶

Write (record): filtered messages are joined into a single ReMeTrajectory and posted to ReMe’s add endpoint; the server then runs LLM extraction over the trajectory to produce searchable memory snippets.
Retrieve: the current message is used as the query against ReMe’s search. The server-aggregated answer field is returned when present, otherwise multiple memory snippets are joined.

Writes use the same filtering as Bailian:

Only USER and ASSISTANT messages are kept.
Assistant messages containing ToolUseBlock (tool-call requests) are skipped.
Messages containing the <compressed_history> marker are skipped.

Builder reference¶

Method	Required	Default	Notes
`userId(String)`	✅	-	Workspace ID (used for both writes and reads)
`apiBaseUrl(String)`	✅	-	ReMe service URL, e.g. `http://localhost:8002`
`timeout(Duration)`	❌	`60s`	HTTP timeout

ReMe does not yet expose finer-grained metadata filtering. If you need tag-based segmentation, encode it inside userId (e.g. tenant-a:project-1).