跳到主要内容
Memory runtime for agents

Memorose turns raw agent traffic into memory that compounds.

It ingests events, consolidates facts and procedures, projects shared organizational knowledge, tracks goals and tasks, and keeps the whole runtime operable through a real dashboard and Raft-backed deployment model.

Start with Quickstart

Read the API

Understand the Model

L0-L3 memory hierarchyagent / user / organizationstreams, tasks, graph, knowledge
POST /eventsPOST /retrieveGET /tasks/treeGET /organizations/:org_id/knowledge
01
Memory runtime

Streams, layered memory, hybrid retrieval, and active forgetting.

02
Execution layer

Task trees, ready queues, milestones, dependencies, and result sedimentation.

03
Operations surface

Dashboard auth, cluster control, organization knowledge, and runtime inspection.

What Ships Now

The docs are organized around the runtime you actually deploy.

This is not a “store one memory, search one memory” product anymore. The pages below are meant to help you reason about the system as a memory runtime, operational surface, and execution layer.

Current API

Stream-first HTTP surface

Ingest and retrieval are scoped by user_id and stream_id, then connected to tasks, graph edges, organization knowledge, and cluster operations.

  • /events for raw ingest
  • /retrieve for hybrid recall
  • /tasks/* for L3 execution state
  • /organizations/:org_id/knowledge for shared knowledge
Inspect the REST surface
Memory model

L0-L3 with clear ownership

Memory hierarchy and memory domains are separate concepts. One describes how memory evolves; the other describes who owns it.

Multimodal

Beyond text-only pipelines

The runtime can ingest and retrieve against text, image, audio, video, and JSON payloads.

Operations

Dashboard and cluster control

Dashboard auth, memory inspection, organization knowledge views, and Raft cluster operations are part of the documented product surface.

Source of truth

Read docs against code, not against old examples

  • README.md
  • crates/memorose-server/src/main.rs
  • config.example.toml
  • .env.example
How To Read It

Follow the same flow the runtime follows.

The fastest way to understand Memorose is to start from runtime motion, not storage internals.

  1. 01

    Ingest events into a stream

    Everything starts as L0 input attached to a concrete user and stream timeline.

  2. 02

    Consolidate into durable memory

    Facts and procedural traces become stable memory units instead of remaining raw traffic forever.

  3. 03

    Retrieve with more than one signal

    Hybrid retrieval combines vector search, text search, graph depth, time filters, and shared organization knowledge.

  4. 04

    Track execution through L3

    Goals, milestones, dependencies, and completion summaries turn memory into an execution surface rather than a passive archive.

Choose By Job

Open the docs from the angle you actually care about.

Build
I need to wire ingest and retrieval into an agent.

Start with Quickstart, REST API, and Hybrid Search.

Operate
I need to inspect memory, tasks, knowledge, and cluster health.

Go through Dashboard, Organization Knowledge, and Distributed Deployment.

Reason
I need the model before I design wrappers or product behavior.

Read Architecture, Memory Hierarchy, Memory Domains, and Forgetting Strategies.

Start Here

Choose the shortest path into the runtime.

Quickstart
Log in, create a stream, ingest an event, retrieve memory.

Use the real v1 routes with the current auth and response model.

Architecture
Understand L0-L3, domains, and the runtime pipeline.

Read the model before you build wrappers around the wrong abstraction.

Tasks
See how goals and execution state are exposed in L3.

Task trees and ready queues are part of the product, not an internal detail.

Organization Knowledge
Inspect the shared memory layer for one org_id.

Use it when personal memory needs to evolve into reusable organizational knowledge.