OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Agent Memory & RAG

Remory

Remory is a Python developer tool that supplements an agent’s compaction summary with learned soft memory derived from its history, then uses the resulting checkpoint to continue generation.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Remory — OpenViST project profile in Agent Memory & RAG.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Remory when an agent must continue a long-running task after conversation history has been compacted and you want to add learned residual memory to the summary.

Who it is for

  • Developers integrating memory into an agent’s context-compaction pipeline.
  • Codex users who want to run the documented Remory compaction integration.
  • Teams deploying the listed base-model and Remory-checkpoint combinations on SGLang or Transformers.

Documented uses

  • Encode information from messages removed during compaction as learned soft memory conditioned on the summary.
  • Return a compacted checkpoint that an agent can use to continue generation.
  • Run Codex sessions with Remory compaction through the documented `r-codex` command.
Before you choose

Requirements & limits

The public GitHub repository declares an MIT license. The documented quick start runs locally with SGLang; it specifies at least two 80 GB GPUs for the Qwen3.8-27B setup.

  • Remory adds summary-conditioned soft memory to compaction; it does not take over the agent’s tools or decide when compaction occurs.
  • The documented quick start combines Qwen3.8-27B with a 1.9B Remory checkpoint and calls for at least two 80 GB GPUs.
  • The README also lists a GLM-5.3-Flash Remory checkpoint for SGLang and Qwen deployment options for SGLang and Transformers.
  • Codex can use the documented integration; research web search and document tools require separate API credentials and a tool environment.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up Remory with the documented Qwen and SGLang quick start

Install the local environment and launch the Qwen3.8-27B model with its Remory checkpoint on SGLang.

Before you start

  • At least two 80 GB GPUs for the documented Qwen3.8-27B plus Remory deployment.
  • The README installer prepares a local environment with a pinned SGLang version, the Remory hook and CUDA compiler components.

Steps

  1. Clone the Remory repository, enter its directory, run the command below, and activate .remory/venv.

    Documented command
    python deploy/install.py

  2. Launch remory.launch_server with the Qwen3.8-27B base model, the Qwen3.8-27B-REMORY-1.9B checkpoint and tensor parallelism set to two.

  3. In another terminal, activate the environment and run r-codex -C /path/to/project to open Codex with Remory compaction.

Expected outcome

A local SGLang service is launched with Remory, and the documented Codex command opens a project session using Remory compaction.

Guide limitations

  • The first run downloads model weights and compiles GPU kernels.
  • The two-GPU requirement applies to the documented Qwen3.8-27B quick start.
  • The optional research tools require Serper and MinerU credentials and a separate Docling/MinerU tool environment.
Documentation 1 ↗
Source-based practical guide

Connect Remory compaction to an agent with Harness

Create a compacted checkpoint from an agent’s retained prefix, removed history and summary instructions, then continue generation from that checkpoint.

Before you start

  • A running Remory inference endpoint.
  • The actor’s native chat-template codec.
  • The agent’s prefix messages, messages to remove, summary prompt and summary schema.

Steps

  1. Create a Remory Client pointed at the inference endpoint and supply a session ID.

  2. Create a Harness using the client and the actor’s native chat-template codec.

  3. Call Harness.compact with the prefix messages, removed messages, summary prompt and summary schema.

  4. Pass the returned checkpoint to Harness.generate with the prefix messages to continue.

Expected outcome

The example integration produces a checkpoint containing a summary and residual memory for continued generation.

Guide limitations

  • The agent retains responsibility for its own tools and compaction trigger.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
23 October 2026
Original website title
fudan-generative-vision/Remory · Official homepage title · retained in its source language
How we review sourcesSuggest a correction