M3-Agent (ByteDance)
https://github.com/ByteDance-Seed/m3-agent
ByteDance Seed + Zhejiang Univ + SJTU. Multimodal agent with long-term memory as entity-centric graph built from live video + audio streams. Parallel memorisation + control processes. Ships M3-Bench evaluation dataset.
At a glance
- Type
- Episodic + semantic multimodal memory graph
- Tier
- T2
- Section
- Dedicated memory layers
- Created
- 2025-07
- Latest release
- no releases
- License
- Apache-2.0
- GitHub
- 1.3k★ Python
- Pricing
- not applicable — academic research code (no commercial pricing)
- Funding
- not applicable — ByteDance corporate research
Taxonomy
- storage
- graph
- retrieval
- graph-traversal
- persistence
- long-term
- update
- extraction
- unit
- episode
- governance
- opaque
- conflict
- n/a
When to use
Optimised for: episodic + semantic multimodal memory graph (research)
Anti-fit: not applicable - research paper
Pros & cons
Pros
Multi-modal multi-memory agent — three coordinated memory streams for video, audio, text agent tasks.
Cons
ByteDance-internal; published architecture but production access limited; eval primarily on Chinese benchmarks.
Claims & capabilities
+6.7% / +7.7% / +5.3% accuracy on M3-Bench-robot / M3-Bench-web / VideoMME-long vs Gemini-1.5-pro + GPT-4o baseline.
Technical surface
- API surface
- searched not found
- Backend storage
- searched not found
- Deployment
- Self-hosted (research code; API config + memory graph files)
- Embedding model
- searched not found
- Multi-tenancy
- searched not found
- MCP
- not documented publicly
- A2A
- not documented publicly
- OpenTelemetry
- not documented publicly
Similar systems
Other dedicated memory layers in the catalog, ranked by inbound references.
- Mem0 T1
Universal memory layer for AI agents. Three concurrent stores (vector + graph + KV); LLM-extracted facts; concurrent retrieval via ThreadPoolExecutor.
- Zep & Graphiti T1
Bi-temporal knowledge graph (event time + ingestion time). Strong on chronological reasoning and contradiction tracking. Graphiti is the open-source core.
- Cognee T1
"Extract–Cognify–Load" pipeline that turns raw input into a typed, queryable knowledge graph for agent recall.
- Hindsight (Vectorize) T1
Standalone memory service from Vectorize. Open source. Biomimetic four-network design (World, Bank, Observation, Opinion). Ships an MCP memory server.
- Memvid T2
Single-file memory layer (one .mv2 file). No DB, no server. Append-only sequence of immutable Smart Frames with timestamps + checksums. Native Rust core (rewritten from Python).
- Supermemory T1
Memory engine with API, app, browser extension, and MCP server. Extracts facts, tracks updates, resolves contradictions, auto-forgets expired info. Plugins for Claude Code, OpenCode, OpenClaw, Hermes.