<p align="center"> <strong>OMEM</strong><br/> Shared Memory That Never Forgets </p> <p align="center"> <a href="https://github.com/ourmem/omem/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-blue.svg" alt="License"></a> <a href="https://ourmem.ai"><img src="https://img.shields.io/badge/hosted-ourmem.ai-green.svg" alt="Hosted"></a> <a href="https://github.com/ourmem/omem"><img src="https://img.shields.io/github/stars/ourmem/omem?style=social" alt="Stars"></a> </p> <p align="center"> <strong>English</strong> | <a href="README_CN.md">简体中文</a> </p>

The Problem

Your AI agents have amnesia — and they work alone.

🧠 Amnesia — every session starts from zero. Preferences, decisions, context — all gone.
🏝️ Silos — your Coder agent can't access what your Writer agent learned.
📁 Local lock-in — memory tied to one machine. Switch devices, lose everything.
🚫 No sharing — team agents can't share what they know. Every agent re-discovers the same things.
🔍 Dumb recall — keyword match only. No semantic understanding, no relevance ranking.
🧩 No collective intelligence — even when agents work on the same team, there's no shared knowledge layer.

ourmem fixes all of this.

What is ourmem

ourmem gives AI agents shared persistent memory — across sessions, devices, agents, and teams. One API key reconnects everything.

🌐 Website: ourmem.ai

🧑‍💻 I use AI coding tools

Install the plugin for your platform. Memory works automatically — your agent recalls past context on session start and captures key info on session end.

→ Jump to Quick Start

</td> <td width="50%" valign="top">

🔧 I'm building AI products

REST API with 48+ endpoints. Docker one-liner for self-deploy. Embed persistent memory into your own agents and workflows.

→ Jump to Self-Deploy

</td> </tr> </table>

Core Capabilities

<table> <tr> <td width="25%" align="center"> <h4>🔗 Shared Across Boundaries</h4> Three-tier Spaces — Personal, Team, Organization — let knowledge flow across agents and teams with full provenance tracking. </td> <td width="25%" align="center"> <h4>🧠 Never Forget</h4> Weibull decay model manages the memory lifecycle — core memories persist, peripheral ones gracefully fade. No manual cleanup. </td> <td width="25%" align="center"> <h4>🔍 Deep Understanding</h4> 11-stage hybrid retrieval: vector search, BM25, RRF fusion, cross-encoder reranking, and MMR diversity for precise recall. </td> <td width="25%" align="center"> <h4>⚡ Smart Evolution</h4> 7-decision reconciliation — CREATE, MERGE, SUPERSEDE, SUPPORT, CONTEXTUALIZE, CONTRADICT, or SKIP — makes memories smarter over time. </td> </tr> </table>

📖 Memory Pipeline Architecture — Technical deep-dive into how ourmem stores, retrieves, and evolves memories.

🔗 Memory Sharing Architecture — How memories flow across agents and teams: sharing, provenance, versioning, and cross-space search.

Feature Overview

| Category | Feature | Details | |----------|---------|---------| | Platforms | 4 platforms | OpenCode, Claude Code, OpenClaw, MCP Server | | Sharing | Space-based sharing | Personal / Team / Organization with provenance | | | Provenance tracking | Every shared memory carries full lineage | | | Quality-gated auto-sharing | Rules fire on memory creation (async, non-blocking) | | | Vector-enabled shared copies | Shared copies carry source vector embeddings for full search | | | Idempotent sharing | Re-sharing returns existing copy (no duplicates) | | | Version tracking | Memories track version counter, shared copies detect staleness via ?check_stale=true | | | Re-share stale copies | Refresh outdated shared copies with latest source content and vector | | | Convenience sharing | One-step cross-user share (share-to-user) and bulk share (share-all-to-user) with auto-bridging | | | Organization management | One-step org creation (org/setup) and publish (org/publish) with auto-share rules | | | Cross-space search | Search across all accessible spaces at once | | Ingestion | Smart dedup | 7 decisions: CREATE, MERGE, SKIP, SUPERSEDE, SUPPORT, CONTEXTUALIZE, CONTRADICT | | | Noise filter | Regex + vector prototypes + feedback learning | | | Admission control | 5-dimension scoring gate (utility, confidence, novelty, recency, type prior) | | | Dual-stream write | Sync fast path (<50ms) + async LLM extraction | | | Post-import intelligence | Batch import → async LLM re-extraction + relation discovery | | | Adaptive import strategy | Auto/atomic/section/document — heuristic content type detection | | | Content fidelity | Original text preserved, dual-path search (vector + BM25 on source text) | | | Cross-reconcile | Discover relations between memories via vector similarity | | | Batch self-dedup | LLM deduplicates facts within same import batch | | | Privacy protection | <private> tag redaction before storage | | Retrieval | 11-stage pipeline | Vector + BM25 → RRF → reranker → decay → importance → MMR diversity | | | User Profile | Static facts + dynamic context, <100ms | | | Retrieval trace | Per-stage explainability (input/output/score/duration) | | Lifecycle | Weibull decay | Tier-specific β (Core=0.8, Working=1.0, Peripheral=1.3) | | | Three-tier promotion | Peripheral ↔ Working ↔ Core with access-based promotion | | | Auto-forgetting | TTL detection for time-sensitive info ("tomorrow", "next week") | | Multi-modal | File processing | PDF, image OCR, video transcription, code AST chunking | | | GitHub connector | Real-time webhook sync for code, issues, PRs | | Deploy | Open source | Apache-2.0 (plugins + docs) | | | Self-hostable | Single binary, Docker one-liner, ~$5/month | | | musl static build | Zero-dependency binary for any Linux x86_64 | | | Object storage | Alibaba Cloud OSS or S3-compatible, with ECS RAM role support | | | Hosted option | ourmem.ai — nothing to deploy |

From Isolated Agents to Collective Intelligence

Most AI memory systems trap knowledge in silos. ourmem's three-tier Space architecture enables knowledge flow across agents and teams — with provenance tracking and quality-gated sharing.

Research shows collaborative memory reduces redundant work by up to 61% — agents stop re-discovering what their teammates already know. — Collaborative Memory, ICLR 2026

| | Personal | Team | Organization | |---|----------|------|--------------| | Scope | One user, multiple agents | Multiple users | Company-wide | | Example | Coder + Writer share preferences | Backend team shares arch decisions | Tech standards, security policies | | Access | Owner's agents only | Team members | All org members (read-only) |

Provenance-tracked sharing — every shared memory carries its lineage: who shared it, when, and where it came from. Shared copies include the source memory's vector embedding, so they're fully searchable in the target space.

Quality-gated auto-sharing — rules filter by importance, category, and tags. Rules fire automatically when new memories are created. Only high-value insights cross space boundaries.

How It Works

┌──────────────────────────────────────────────────────────────────┐
│   Your AI Agent (OpenCode / Claude Code / OpenClaw / Cursor)     │
│                                                                  │
│   Session Start → auto-recall relevant memories                  │
│   During Work   → keyword detection triggers recall              │
│   Session End   → auto-capture decisions, preferences, facts     │
└───────────────────────────┬──────────────────────────────────────┘
                            │ REST API (X-API-Key)
                            ▼
┌──────────────────────────────────────────────────────────────────┐
│                       ourmem Server                              │
│                                                                  │
│  ┌─ Smart Ingest ─────────────────────────────────────────────┐  │
│  │  Messages → LLM extraction → noise filter → admission      │  │
│  │  → 7-decision reconciliation (CREATE / MERGE / SUPERSEDE / │  │
│  │    SUPPORT / CONTEXTUALIZE / CONTRADICT / SKIP)            │  │
│  │  → cross-reconcile relations → privacy redaction           │  │
│  └────────────────────────────────────────────────────────────┘  │
│                                                                  │
│  ┌─ Hybrid Search (11 stages) ────────────────────────────────┐  │
│  │  Vector + BM25 → RRF fusion → cross-encoder reranker       │  │
│  │  → Weibull decay boost → importance scoring                │  │
│  │  → MMR diversity → parallel cross-space aggregation        │  │
│  └────────────────────────────────────────────────────────────┘  │
│                                                                  │
│  ┌─ Sharing Engine ───────────────────────────────────────────┐  │
│  │  Personal / Team / Organization spaces                     │  │
│  │  → provenance tracking → version-based stale detection     │  │
│  │  → auto-share rules → one-step share-to-user              │  │
│  └────────────────────────────────────────────────────────────┘  │
│                                                                  │
│  ┌─ Lifecycle ────────────────────────────────────────────────┐  │
│  │  Weibull decay (Core β=0.8 / Working β=1.0 / Peripheral   │  │
│  │  β=1.3) → 3-tier promotion → auto-forgetting TTL          │  │
│  └────────────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────────────┘

Write once, recall everywhere — memories persist across sessions, devices, and agents
Gets smarter over time — reconciliation merges, updates, and contradicts memories automatically
Share across boundaries — Personal → Team → Or

Omem

Install / Use

README