> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentscope.io/llms.txt
> Use this file to discover all available pages before exploring further.
> ## Agent Instructions
> For AgentScope Python, use https://docs.agentscope.io/stable/en/index for new projects. For existing projects, check the installed agentscope version and use matching versioned documentation.
> The /latest/ alias points to development documentation. Use it only with the matching development source. Do not mix AgentScope 1.x and 2.x APIs.
> State the AgentScope version when providing installation commands or code examples. ReMe uses its own continuously updated /reme/latest/ documentation.
# What's ReMe?
> Learn how ReMe turns conversations and resources into user-owned, file-native agent memory.
ReMe is a local-first, self-evolving personal knowledge base for AI agents. It turns conversations and resources into readable, editable, searchable, and interconnected Markdown memory, then keeps that memory organized for future recall. The files remain under the user's control and can be shared across agents such as QwenPaw, OpenClaw, Hermes, Claude Code, and Codex.
The following diagram summarizes ReMe's file-native approach to shared, portable memory:
Install ReMe and complete a verifiable write, search, and read loop.
Understand the principles and trade-offs behind file-native memory.
Embed ReMe in Python, connect a supported agent, or choose a service interface.
## Core Ideas
* **Memory as File, File as Memory**: memory is a user-owned file, not hidden database state; users can inspect, edit, move, and back it up directly.
* **Memory from Experience**: long-term memory is refined and corrected, not an endless context dump.
* **Human-Agent Shared Memory**: people and agents work on the same visible, portable Markdown.
* **Connected and Traceable**: Auto Link writes relationships and sources back into the files, providing context and tracing durable conclusions to evidence.
Read the [core concepts](/en/reme/latest/concepts) for the full philosophy and trade-offs.
## What ReMe Solves
Agent context usually disappears when a session ends. Even with a vector database, users may not know what was remembered, where a conclusion came from, or how to correct and move that memory.
ReMe provides a complete memory loop in a local workspace without hiding the source data:
* **Capture**: raw conversations and resources remain available as traceable sources.
* **Refine**: useful facts, preferences, procedures, and knowledge evolve into Markdown memory.
* **Retrieve**: BM25 and bounded wikilink expansion work by default; embeddings add optional semantic recall and reciprocal rank fusion.
* **Integrate**: Python, AgentScope, MCP, CLI, HTTP, ReMe Studio, plugins, and Skills expose the same file-native memory model; keep one runtime owner per workspace.
Memory becomes shared knowledge owned by the user instead of opaque agent state.
## Where It Fits
* **Personal assistants** that retain preferences, background, plans, and important events.
* **Coding agents** that carry project decisions, conventions, and debugging experience across sessions.
* **Personal knowledge bases** built from conversations, notes, and linked resources.
* **Experience-driven agents** that preserve reusable procedures and lessons learned.
ReMe may need another system alongside it when you need strict transactions, complex authorization, or very large-scale online retrieval. It is a memory layer—not a complete agent product—and deliberately avoids invisible, unbounded collection.
## File-Native Memory System
ReMe separates raw evidence, intermediate refinement, durable memory, and derived system state. The diagram shows how people and agents work with the same files while rebuildable indexes support retrieval.
The workspace maps the flow to these top-level directories:
| Layer | Directories | Role |
| -------------- | ----------------------- | -------------------------------------------------------------------------------- |
| Sources | `session/`, `resource/` | Preserve source conversations and imported material |
| Agent runtime | `mem_session/` | Hold generated agent-wrapper sessions and configuration; this is not user memory |
| Daily memory | `daily/` | Hold source-linked summaries and lightly refined memory |
| Durable memory | `digest/` | Store reusable personal facts, procedures, and knowledge |
| Derived state | `metadata/` | Maintain rebuildable indexes, graphs, catalogs, and caches |
Auto Memory turns conversations into topic-named daily cards, Auto Resource interprets files placed at the root of `resource/` or under a dated subdirectory, and Auto Dream consolidates changed daily cards into durable knowledge. By default, Auto Dream examines the latest two-day window and extracts at most five reusable units before creating, corroborating, refining, or correcting digest nodes. The default live watcher indexes Markdown under `daily/` and `digest/`; a full `reme reindex` also scans `resource/` and JSONL. Search returns matching chunks with paths and line ranges, followed by bounded incoming and outgoing wikilink metadata so agents can expand context only when needed.
`proactive` only reads topics produced by Auto Dream. It does not browse the web, send notifications, or rewrite memory on its own; the host agent decides whether and how to act.
## Offline Use
File operations, keyword search, wikilink traversal, and reading proactive topics do not require an LLM. Conversation refinement, resource interpretation, and long-term consolidation need a compatible LLM. Semantic search additionally needs an embedding service.
## Choose an Integration
Embed ReMe directly or use AgentScope's first-party `ReMeMiddleware`.
Connect service-based agents such as Claude Code, Hermes, OpenClaw, or Codex.
Connect through MCP, a Skill, the CLI, or HTTP.
## Published Benchmark Results
Repository reference runs report 89.4% over 500 LongMemEval cleaned-s questions, 66.1% on the 100K BEAM setting, 65.0% on the 1M BEAM setting, and a 0.580 PROC score across five personas in π-Bench. See the [ReMe benchmark directories](https://github.com/agentscope-ai/ReMe/tree/main/benchmark) for model, prompt, dataset, and judging details; compare results only under matching evaluation settings.
## Research
ReMe's experience-driven memory direction is described in [Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution](https://aclanthology.org/2026.findings-acl.829/), accepted to Findings of ACL 2026.
## More
Workspace, configuration, and self-describing interfaces.
Common questions and troubleshooting.
Development setup, tests, and contribution flow.