memos-os 2.2.0

Memory Operating System for LLM Agents — persistent, smart, local-first memory

MemOS — Memory Operating System for AI Agents

Persistent, structured, self-maintaining memory for any LLM agent. Local-first. Framework-agnostic. Connects via MCP to Claude Code, OpenClaw, Cursor, or any HTTP client.

---

What's new in v1.1.0

• Security hardening — WebSocket auth, CORS defaults, Pydantic request schemas on all endpoints
• Canvas force-graph dashboard — clustering, depth filter, time-lapse slider, KG edges
• Modular frontend — monolithic 1768L split into 12 JS modules
• 1710 tests — modernized with shared fixtures, freezegun, tmp_path
• Zero ruff errors — 506 lint errors fixed, full formatting pass
• CI: Python 3.11 / 3.12 / 3.13 + coverage via Codecov

See CHANGELOG.md for full history.

---

Installation

pip install memos-os

With local semantic recall (no external services):

pip install "memos-os[local]"    # sentence-transformers, backend="local"

With vector backend (recommended for production):

pip install "memos-os[chroma]"   # ChromaDB + Ollama embeddings
pip install "memos-os[qdrant]"   # Qdrant
pip install "memos-os[all]"      # all backends

---

Quick start

# Store a memory
memos learn "FastAPI is better than Flask for async workloads" --tags python,backend

# Search semantically
memos recall "which web framework should I use?"

# Start the REST API + dashboard
memos serve --port 8100
# → http://localhost:8100/dashboard

---

Golden path

The full lifecycle from storing a memory to maintaining it over time.

from memos import MemOS
from memos.context import ContextStack
from memos.kg_bridge import KGBridge

mem = MemOS(backend="chroma", embed_host="http://localhost:11434")
cs = ContextStack(mem)
bridge = KGBridge(mem)

# 1. Learn — store memories with tags and importance
mem.learn("User prefers dark mode in all apps", tags=["preference", "ui"], importance=0.8)
mem.learn("Deploy with Docker on ARM64 homelab", tags=["devops", "docker"], importance=0.6)
mem.learn("Alice leads the backend team", tags=["team", "people"], importance=0.7)

# 2. Recall — semantic search
results = mem.recall("who handles the server side?", top=3)
for r in results:
    print(f"[{r.score:.2f}] {r.item.content}")
# [0.91] Alice leads the backend team (tags: team, people)

# 3. context_for — targeted context for a specific LLM call
ctx = cs.context_for("how should I deploy this app?", max_chars=1500, top=5)
# Returns a string with identity + relevant memories:
# === RELEVANT MEMORIES (2 results for: 'how should I deploy this app?') ===
# [0.87] Deploy with Docker on ARM64 homelab (tags: devops, docker)

# 4. wake_up — inject at session start
prompt_fragment = cs.wake_up(max_chars=2000, l1_top=15, include_stats=True)
# Returns L0 identity + L1 top-importance memories as a string
# ready to paste into a system prompt

# 5. Reinforce and decay — keep memories healthy over time
# Boost a memory that keeps being useful
results = mem.recall("deployment strategy", top=1)
if results:
    mem._decay.reinforce(results[0].item, strength=0.1)
    mem._store.upsert(results[0].item, namespace=mem._namespace)

# Decay stale memories (preview first)
items = mem._store.list_all(namespace=mem._namespace)
report = mem._decay.run_decay(items, dry_run=True)
print(f"{report.decayed}/{report.total} memories would decay")
# When ready: run_decay(items, dry_run=False)

---

Python SDK

from memos import MemOS

# In-memory (zero dependencies, great for testing)
mem = MemOS()

# JSON persistence
mem = MemOS(backend="json", persist_path="~/.memos/store.json")

# Local-first semantic recall, no Ollama/Chroma required
mem = MemOS(backend="local", persist_path="~/.memos/store.json")

# ChromaDB with local Ollama embeddings
mem = MemOS(backend="chroma", embed_host="http://localhost:11434")

# Qdrant
mem = MemOS(backend="qdrant", qdrant_path="/data/memos")

# Store
mem.learn("User prefers dark mode", tags=["preference", "ui"], importance=0.8)

# Recall
results = mem.recall("what does the user like?", top=5)
for r in results:
    print(f"[{r.score:.2f}] {r.item.content}")

# Forget
mem.forget("memory-id")           # by id
mem.delete_tag("old-project")     # all memories with this tag
mem.prune(threshold=0.2)          # decay-based cleanup

# Stats
s = mem.stats()
# MemoryStats(total_memories=142, avg_relevance=0.71, decay_candidates=8)

---

Which recall API should I use?

Function	Best for	Returns	When NOT to use
mem.recall()	General search, browsing results	list[RecallResult] with scores	You need a ready-made prompt string
memory_search (MCP)	Same as recall(), but over MCP	JSON via MCP protocol	You're using the Python SDK directly
context_for()	Augmenting a single LLM call	str (identity + top results)	You need structured data to process
recall_enriched()	Answers needing entity context	dict with memories + KG facts	No KG data exists or entity resolution isn't needed

from memos.context import ContextStack
from memos.kg_bridge import KGBridge

cs = ContextStack(mem)
bridge = KGBridge(mem)

# recall() — structured results you iterate over
results = mem.recall("docker deployment", top=5)
for r in results:
    print(r.item.content, r.score)

# context_for() — one string, ready for a system prompt
ctx = cs.context_for("docker deployment", max_chars=1000, top=5)

# recall_enriched() — memories + knowledge graph facts in one dict
enriched = bridge.recall_enriched("who is Alice?", top=5, min_score=0.3)
print(enriched["facts"])       # KG triples about Alice
print(enriched["memory_count"])

---

MCP — connect any agent

MemOS exposes a universal MCP endpoint. Any agent that speaks MCP can use it without any code changes.

HTTP (recommended)

Claude Code — add to ~/.claude.json:

{
  "mcpServers": {
    "memos": { "type": "http", "url": "http://localhost:8100/mcp" }
  }
}

OpenClaw — add to ~/.openclaw/openclaw.json:

{
  "mcp": {
    "servers": {
      "memos": { "type": "http", "url": "http://localhost:8100/mcp" }
    }
  }
}

Any MCP client — POST http://localhost:8100/mcp with JSON-RPC 2.0 body.

Discovery: GET http://localhost:8100/.well-known/mcp.json

Stdio (Claude Code local)

{
  "mcpServers": {
    "memos": { "command": "memos", "args": ["mcp-stdio"] }
  }
}

Or run standalone: memos mcp-serve --port 8200

Available MCP tools

Tool	Description
memory_search	Semantic search — query, top_k, tags
memory_save	Store a memory — content, tags, importance
memory_forget	Delete by id or tag
memory_stats	Counts, avg importance, decay candidates
memory_wake_up	Identity + top memories ready to inject at session start
memory_context_for	Context optimised for a specific query
memory_decay	Run decay cycle (dry-run by default)
memory_reinforce	Boost a memory's importance score
kg_add_fact	Add a temporal triple to the Knowledge Graph
kg_query_entity	All active facts for an entity
kg_timeline	Chronological fact history for an entity
memory_recall_enriched	Memories + KG facts in one call

---

REST API

Start the server: memos serve --port 8100

Interactive docs: http://localhost:8100/docs

POST   /api/v1/learn                Store a memory
POST   /api/v1/learn/batch          Bulk store
POST   /api/v1/recall               Semantic search
GET    /api/v1/recall/stream        SSE streaming recall
GET    /api/v1/search               Keyword search
GET    /api/v1/stats                Memory statistics
GET    /api/v1/tags                 List all tags
DELETE /api/v1/tags/{tag}           Delete tag from all memories
DELETE /api/v1/memory/{id}          Delete a memory
POST   /api/v1/prune                Decay-based cleanup
GET    /api/v1/graph                Knowledge graph (nodes + edges for D3.js)
GET    /api/v1/export/parquet       Download .parquet backup
POST   /mcp                         MCP JSON-RPC endpoint
GET    /.well-known/mcp.json        MCP discovery
GET    /dashboard                   Second Brain UI
GET    /health                      Health check

---

Configuration

All options can be set via environment variables:

MEMOS_BACKEND=chroma              # memory | json | chroma | qdrant | pinecone
MEMOS_NAMESPACE=default           # memory namespace (one per agent)
MEMOS_PERSIST_PATH=~/.memos/      # path for json/sqlite storage

# ChromaDB
MEMOS_CHROMA_URL=http://chroma:8000
MEMOS_EMBED_HOST=http://localhost:11434   # Ollama — bypasses server-side ONNX
MEMOS_EMBED_MODEL=nomic-embed-text

# Qdrant
MEMOS_QDRANT_HOST=localhost
MEMOS_QDRANT_PORT=6333

# Pinecone
MEMOS_PINECONE_API_KEY=***
MEMOS_PINECONE_INDEX=agent-memories

---

Docker

Single container (JSON backend, no dependencies):

docker run -p 8100:8000 \
  -e MEMOS_BACKEND=json \
  -v memos-data:/root/.memos \
  ghcr.io/mars375/memos:latest

Full stack with ChromaDB + Ollama embeddings:

git clone https://github.com/Mars375/memos
cd memos
docker compose up -d

---

Import conversations

Mine your existing conversations into MemOS:

# Auto-detect format
memos mine conversations.json

# Supported formats
memos mine export.json      --format claude      # Claude Projects export
memos mine conversations.json --format chatgpt   # ChatGPT export
memos mine messages.json    --format discord
memos mine result.json      --format telegram    # Telegram Desktop export
memos mine channel.jsonl    --format slack
memos mine ~/.openclaw/workspace-labs/ --format openclaw

# Options
memos mine ~/notes/ --dry-run --tags project-x --chunk-size 600 --namespace agent-alice

Python API:

from memos.ingest.miner import Miner

miner = Miner(mem, chunk_size=800, chunk_overlap=100)
result = miner.mine_auto("conversations/")   # auto-detect
# MineResult(imported=127, dupes=12, empty=3, errors=0)

---

Knowledge Graph

Store and query temporal facts between entities:

memos kg-add "Alice" "works-at" "Acme Corp" --from 2024-01-01
memos kg-query Alice
memos kg-path Alice Carol --max-hops 3
memos kg-neighbors Alice --depth 2
memos kg-timeline Alice

from memos.knowledge_graph import KnowledgeGraph

kg = KnowledgeGraph()
kg.add_fact("Alice", "works-at", "Acme Corp", valid_from="2024-01-01")
facts = kg.query("Alice")
paths = kg.find_paths("Alice", "Carol", max_hops=3)

---

Living Wiki

Compile memories into entity-based markdown pages with backlinks:

memos wiki-living update          # scan memories, create/update entity pages
memos wiki-living read Alice      # print Alice's page
memos wiki-living search "python" # search across all pages
memos wiki-living lint            # find orphans, contradictions, empty pages

---

Memory decay

Memories age automatically. Important ones persist; stale ones fade.

memos decay --dry-run      # preview what would decay
memos decay --apply        # apply decay
memos prune --threshold 0.1  # delete memories below importance threshold

---

Versioning and time-travel

Every write is versioned. Query the past, diff changes, roll back.

memos history <memory-id>
memos diff <memory-id> --latest
memos rollback <memory-id> --version 1 --yes
memos recall-at "user preferences" --at 2d    # as of 2 days ago
memos snapshot-at 1w                          # all memories 1 week ago

---

Multi-namespace (multi-agent)

Each agent gets its own isolated namespace:

memos --namespace agent-alice learn "Alice's memory"
memos --namespace agent-bob learn "Bob's memory"
memos --namespace agent-alice recall "what do I know?"

---

Development

git clone https://github.com/Mars375/memos
cd memos
pip install -e ".[dev]"

# Lint
ruff check src/ tests/
ruff format src/ tests/

# Tests (Python 3.11 / 3.12 / 3.13)
pytest -q --tb=short          # 1710 tests
pytest tests/test_core.py     # specific module

---

Architecture

MemOS is built around three core layers:

• Capture — Mine conversations and events into structured memory units via the CLI, SDK, or MCP.
• Engine — Storage, recall, decay, reinforcement, versioning, and knowledge graph. Pluggable backends (in-memory, JSON, ChromaDB, Qdrant, Pinecone).
• Knowledge Surface — Living wiki, graph view, and context packs (wake_up, context_for, recall_enriched) that serve the right context at the right time.

See ROADMAP.md for planned features and PRD.md for product requirements.

---

License

MIT — Mars375