OMEM
Shared Memory That Never Forgets

English | 简体中文

--- ## 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](https://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](#quick-start)**

### 🔧 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](#self-deploy)**

## Core Capabilities

🔗 Shared Across Boundaries

Three-tier Spaces — Personal, Team, Organization — let knowledge flow across agents and teams with full provenance tracking.

🧠 Never Forget

Weibull decay model manages the memory lifecycle — core memories persist, peripheral ones gracefully fade. No manual cleanup.

🔍 Deep Understanding

11-stage hybrid retrieval: vector search, BM25, RRF fusion, cross-encoder reranking, and MMR diversity for precise recall.

⚡ Smart Evolution

7-decision reconciliation — CREATE, MERGE, SUPERSEDE, SUPPORT, CONTEXTUALIZE, CONTRADICT, or SKIP — makes memories smarter over time.

📖 **[Memory Pipeline Architecture](docs/PIPELINE.md)** — Technical deep-dive into how ourmem stores, retrieves, and evolves memories. 🔗 **[Memory Sharing Architecture](docs/SHARING.md)** — 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 | `` 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 | | | Self-hostable | Single binary, Docker one-liner, ~$5/month | | | musl static build | Zero-dependency binary for any Linux x86_64 | | | Object storage | AWS S3 or any S3-compatible storage, with IAM 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 → Organization knowledge flow with full provenance - **No manual memory management** — Weibull decay handles lifecycle, auto-share rules handle distribution ## Key Concepts ### API Key = Tenant ID = Your Identity When you create a tenant, the returned `id` **is** your API Key. They're the same UUID. There is no separate "tenant ID". ```bash curl -X POST https://api.ourmem.ai/v1/tenants -d '{"name": "alice"}' # → {"id": "a1b2c3d4-...", "api_key": "a1b2c3d4-...", "status": "active"} # ↑ same value ↑ ``` ### API Key vs Space ID | Concept | What it is | Example | How many | |---------|-----------|---------|----------| | **API Key** | Your identity. Goes in `X-API-Key` header. | `a1b2c3d4-...` | 1 per user | | **Space ID** | A memory storage address. Each is an isolated database. | `personal/a1b2c3d4-...` | Multiple per user | One API Key owns multiple Spaces: ``` API Key "a1b2c3d4" │ ├── personal/a1b2c3d4 ← auto-created, your private memories ├── team/e5f6g7h8 ← team space you created (you = Admin) ├── team/i9j0k1l2 ← team space you were invited to (you = Member) └── org/m3n4o5p6 ← organization you joined (you = Reader) ``` ### Sharing = Passing Your API Key To share memories with another user, you pass their API Key as `target_user`: ```bash # Share all your preferences to user Bob curl -X POST https://api.ourmem.ai/v1/memories/share-all-to-user \ -H "X-API-Key: your-api-key" \ -d '{"target_user": "bobs-api-key", "filters": {"categories": ["preference"]}}' ``` The system automatically creates a bridging Team Space between you and Bob. No manual space management needed. ## Quick Start ### Agent Install (recommended) One message to your AI agent. It handles everything — API key, plugin install, config, verification. **Hosted (ourmem.ai — nothing to deploy):** | Platform | Copy this to your agent | |----------|------------------------| | **OpenClaw** | `Read https://ourmem.ai/SKILL.md and follow the instructions to install and configure ourmem for OpenClaw` | | **Claude Code** | `Read https://ourmem.ai/SKILL.md and follow the instructions to install and configure ourmem for Claude Code` | | **OpenCode** | `Read https://ourmem.ai/SKILL.md and follow the instructions to install and configure ourmem for OpenCode` | | **Cursor / VS Code** | `Read https://ourmem.ai/SKILL.md and follow the instructions to install and configure ourmem as MCP Server` | **Self-hosted (your own server):** | Platform | How to install | |----------|---------------| | **OpenClaw** | Run `openclaw skills install ourmem`, then tell your agent: `setup ourmem in self-hosted mode` | | **Claude Code** | `Read https://raw.githubusercontent.com/ourmem/omem/main/skills/ourmem/SKILL.md and install ourmem for Claude Code, self-hosted mode` | | **OpenCode** | `Read https://raw.githubusercontent.com/ourmem/omem/main/skills/ourmem/SKILL.md and install ourmem for OpenCode, self-hosted mode` | That's it. Your agent handles the rest. **Skill Install (alternative):** If you prefer CLI installation, install the ourmem skill directly: ```bash npx skills add ourmem/omem --skill ourmem -g ``` This works with 44+ AI agents including Claude Code, OpenCode, Cursor, and more. See [Vercel Skills CLI](https://github.com/vercel-labs/skills) for details. ---

Manual Install (without agent assistance)

### 1. Get an API Key **Hosted:** ```bash curl -sX POST https://api.ourmem.ai/v1/tenants \ -H "Content-Type: application/json" \ -d '{"name": "my-workspace"}' | jq . # → {"id": "xxx", "api_key": "xxx", "status": "active"} ``` **Self-deploy:** ```bash docker run -d -p 8080:8080 -e OMEM_EMBED_PROVIDER=bedrock ghcr.io/ourmem/omem-server:latest curl -sX POST http://localhost:8080/v1/tenants \ -H "Content-Type: application/json" \ -d '{"name": "my-workspace"}' | jq . ``` Save the returned `api_key` — this reconnects you to the same memory from any machine. ### 2. Install Plugin **OpenCode:** Add `"plugin": ["@ourmem/opencode"]` to `opencode.json` + configure `plugin_config` with `apiUrl` and `apiKey` in the same file. **Claude Code:** `/plugin marketplace add ourmem/omem` + set env vars in `~/.claude/settings.json`. **OpenClaw:** `openclaw plugins install @ourmem/ourmem` + configure `openclaw.json` with apiUrl and apiKey. **MCP (Cursor / VS Code / Claude Desktop):** ```json { "mcpServers": { "ourmem": { "command": "npx", "args": ["-y", "@ourmem/mcp"], "env": { "OMEM_API_URL": "https://api.ourmem.ai", "OMEM_API_KEY": "your-api-key" } } } } ``` ### 3. Verify ```bash curl -sX POST "$OMEM_API_URL/v1/memories" \ -H "X-API-Key: $OMEM_API_KEY" -H "Content-Type: application/json" \ -d '{"content": "I prefer dark mode", "tags": ["preference"]}' curl -s "$OMEM_API_URL/v1/memories/search?q=dark+mode" -H "X-API-Key: $OMEM_API_KEY" ```

## What Your Agent Gets ### Memory Tools | Tool | Purpose | |------|---------| | `memory_store` | Save facts, decisions, preferences with smart dedup | | `memory_search` | 11-stage hybrid search (vector + BM25 + reranker) | | `memory_get` | Retrieve a specific memory by ID | | `memory_update` | Modify content, tags, importance, tier | | `memory_delete` | Remove a memory | | `memory_list` | Browse memories with filters and pagination | | `memory_ingest` | Smart-ingest full conversations (LLM extraction) | | `memory_profile` | Auto-generated user profile (static facts + dynamic context) | | `memory_stats` | Analytics: tag distribution, decay curves, relation graphs | ### Sharing Tools | Tool | Purpose | |------|---------| | `space_create` | Create Team or Organization spaces | | `space_list` | List all accessible spaces | | `space_add_member` | Invite users to a shared space | | `memory_share` | Share a memory to any space with provenance | | `memory_pull` | Pull a shared memory into your personal space | | `memory_reshare` | Refresh stale shared copies with latest content | > **Convenience APIs:** `share-to-user` (one-step cross-user share with auto-bridging) and `share-all-to-user` (bulk share). MCP exposes 15 tools + 1 resource. OpenCode/OpenClaw expose 11 tools + 3 hooks. ### Smart Hooks | Hook | Trigger | What Happens | |------|---------|--------------| | **SessionStart** | New session begins | Recent memories + user profile auto-injected into context | | **KeywordRecall** | Keyword detected mid-session | Relevant memories surfaced without explicit search | | **SessionEnd** | Session ends | Key decisions, preferences, and facts auto-captured | | **Compaction** | Context window compressed | Critical memories preserved across compaction boundaries | ## Memory Space Browse, search, and manage your agent's memories visually at **[ourmem.ai/space](https://ourmem.ai/space)** — see how memories connect, evolve, and decay over time. ## Security & Privacy | | | |---|---| | **Rust Memory Safety** | No garbage collector, no data races. Ownership model guarantees safety at compile time. | | **Tenant Isolation** | X-API-Key auth with query-level tenant filtering. Every operation verifies ownership. | | **Privacy Protection** | `` tag redaction strips sensitive content before storage. | | **Encryption** | HTTPS for all API transit. Server-side encryption at rest on S3. | | **Admission Control** | 5-dimension scoring gate rejects low-quality data before storage. | | **Open Source Auditable** | Apache-2.0 licensed. Audit every line, fork it, run your own instance. | ## Self-Deploy ```bash # Minimal (BM25 search only, no embedding API needed) docker run -d -p 8080:8080 ghcr.io/ourmem/omem-server:latest # With Bedrock embedding (recommended, needs AWS credentials) docker run -d -p 8080:8080 \ -e OMEM_EMBED_PROVIDER=bedrock \ -e AWS_REGION=us-east-1 \ ghcr.io/ourmem/omem-server:latest # With OpenAI-compatible embedding docker run -d -p 8080:8080 \ -e OMEM_EMBED_PROVIDER=openai-compatible \ -e OMEM_EMBED_API_KEY=sk-xxx \ ghcr.io/ourmem/omem-server:latest ``` Full deployment guide: [docs/DEPLOY.md](docs/DEPLOY.md) ### Object Storage (Optional) By default ourmem stores data on local disk. For durability and scalability, configure AWS S3 or any S3-compatible object storage: ```bash # Add to your environment OMEM_S3_BUCKET=your-bucket # enables s3:// scheme AWS_ENDPOINT_URL=https://s3.amazonaws.com AWS_REGION=us-east-1 # Alternative: OSS-compatible storage OMEM_OSS_BUCKET=your-bucket # enables oss:// scheme OSS_ENDPOINT=https://oss-xx-internal.aliyuncs.com OSS_ACCESS_KEY_ID=your-ak # or use ECS RAM role (auto-discovered) OSS_ACCESS_KEY_SECRET=your-sk ``` > If both `OMEM_S3_BUCKET` and `OMEM_OSS_BUCKET` are set, OSS takes priority. ## Build from Source ### Two build modes | Mode | Command | Binary | Bedrock | Runs on | |------|---------|--------|---------|---------| | **glibc (full)** | `cargo build --release` | Dynamic linked, ~218MB | ✅ AWS Bedrock | Same glibc version as build host | | **musl (portable)** | See below | Static linked, ~182MB | ❌ OpenAI-compatible only | **Any Linux x86_64** | ### glibc build (with Bedrock support) ```bash cargo build --release -p omem-server # Binary: target/release/omem-server # Requires: same or newer glibc on target machine ``` ### musl static build (portable, zero dependencies) Single binary that runs on **any Linux x86_64** — no glibc, no libraries, nothing. ```bash rustup target add x86_64-unknown-linux-musl RUSTFLAGS="-C target-feature=+crt-static -C relocation-model=static" \ cargo build --release --target x86_64-unknown-linux-musl \ -p omem-server --no-default-features # Binary: target/x86_64-unknown-linux-musl/release/omem-server # Statically linked, runs anywhere ``` > **Note:** The musl build uses `--no-default-features` which excludes AWS Bedrock support. Use `OMEM_EMBED_PROVIDER=openai-compatible` (e.g. DashScope, OpenAI) instead. This is because `aws-lc-sys` (AWS crypto library) crashes on musl static linking due to `dlopen(NULL)` incompatibility ([aws-c-cal#213](https://github.com/awslabs/aws-c-cal/issues/213)), and Rust's default `static-pie` output segfaults with musl-gcc ([rust-lang/rust#95926](https://github.com/rust-lang/rust/issues/95926)). ### Transfer to any server ```bash # Compress gzip -c target/x86_64-unknown-linux-musl/release/omem-server > omem-server.gz # Copy to server scp omem-server.gz user@server:/opt/ # Run (no dependencies needed) ssh user@server "gunzip /opt/omem-server.gz && chmod +x /opt/omem-server && /opt/omem-server" ``` ## API at a Glance | Method | Endpoint | Description | |--------|----------|-------------| | POST | `/v1/tenants` | Create workspace & get API key | | POST | `/v1/memories` | Store memory or smart-ingest conversation | | GET | `/v1/memories/search` | 11-stage hybrid search | | GET | `/v1/memories` | List with filters & pagination | | GET | `/v1/profile` | Auto-generated user profile | | POST | `/v1/spaces` | Create shared space | | POST | `/v1/memories/:id/share` | Share memory to a space | | POST | `/v1/files` | Upload PDF / image / video / code | | GET | `/v1/stats` | Analytics & insights | Full API reference (48+ endpoints): [docs/API.md](docs/API.md) ## Documentation | Document | Description | |----------|-------------| | [docs/API.md](docs/API.md) | Complete REST API reference | | [docs/PIPELINE.md](docs/PIPELINE.md) | Memory pipeline architecture — storage, retrieval, and plugin integration flows | | [docs/SHARING.md](docs/SHARING.md) | Memory sharing architecture, flows, and tutorials | | [docs/DEPLOY.md](docs/DEPLOY.md) | Docker & AWS deployment guide | | [docs/PLUGINS.md](docs/PLUGINS.md) | Plugin installation for all 4 platforms | | [docs/TWEETCLAW_OPENCLAW.md](docs/TWEETCLAW_OPENCLAW.md) | OpenClaw workflow for storing TweetClaw X/Twitter research decisions in ourmem | | [skills/ourmem/SKILL.md](skills/ourmem/SKILL.md) | AI agent onboarding skill | ## License Apache-2.0 ---

Shared Memory That Never Forgets.
ourmem.ai · GitHub