Titans (Google)
https://research.google/blog/titans-miras-helping-ai-have-long-term-memory/
Neural long-term memory module that learns to memorise at test time. Uses gradient-of-loss as "surprise" signal; adaptive weight-decay forgetting. Three variants: MAC (memory-as-context), MAG (memory-as-gate), MAL (memory-as-layer).
At a glance
- Type
- Neural fast/slow + surprise gating
- Tier
- T4
- Section
- Research / specialised systems
- Created
- 2025-01
- Latest release
- not applicable — not OSS
- License
- not applicable — not OSS
- GitHub
- not applicable — no GitHub repo
- Pricing
- not applicable — not commercial
- Funding
- not applicable — not commercial
Taxonomy
- storage
- parametric
- retrieval
- attention
- persistence
- session
- update
- parametric-edit
- unit
- kv-token
- governance
- opaque
- conflict
- n/a
When to use
Optimised for: research positioning (see memory_model)
Anti-fit: not applicable - research paper
Pros & cons
Pros
Google's response to Mamba — long-term memory module learns at test time; widely discussed in 2025.
Cons
Closed weights; reproducibility limited; benchmark coverage narrower than published claims suggest.
Claims & capabilities
Scales to 2M+ tokens. Reports outperforming Transformers + recent linear-recurrent models on language modeling, common-sense reasoning, genomics, and time series.
Technical surface
- API surface
- not applicable — research paper
- Backend storage
- not applicable — research paper
- Deployment
- not applicable — not a deployable product
- Embedding model
- not applicable — research paper
- Multi-tenancy
- not applicable — research paper
- MCP
- not applicable — research paper, no deployed product
- A2A
- not applicable — research paper, no deployed product
- OpenTelemetry
- not applicable — research paper, no deployed product
Similar systems
Other research / specialised systems in the catalog, ranked by inbound references.
- A-MEM T3
Treats memories as atomic linkable notes — explicit nod to Zettelkasten knowledge management. Dynamic linking; retroactive memory revision.
- BAI-LAB MemoryOS T3
Hierarchical "OS" with Storage / Updating / Retrieval / Generation modules. Short-term → mid-term via FIFO dialogue-chain; mid-term → long-term via segmented paging.
- EverMemOS T4
Self-organizing memory OS for structured long-horizon reasoning. Three-phase model: episodic, semantic, reconstructive.
- EVOLVE-MEM T3
Dynamic Memory Network + Hierarchical Memory Manager + Self-Improvement Engine. L0 raw embeddings, L1 contextual summaries, L2 high-level principles. NeurIPS 2025 (Scaling Environments for Agents workshop).
- LiCoMemory T4
Lightweight hierarchical graph (CogniGraph) with entities and relations as semantic indexing layers. Incremental graph construction, fast updates, low-latency inference. Nov 2025.
- MemoRAG T3
RAG framework on top of a long-context memory model. Builds global memory once, generates contextual clues at query time. TheWebConf 2025.
Related systems
References (5)
- BABILong cites — S2 isInfluential citation
- Cartesia Sonic cites — S2 isInfluential citation
- Mamba-2 / SSD cites — S2 isInfluential citation
- Recurrent Memory Transformer (RMT) cites — S2 isInfluential citation
- Test-time training (TTT) cites — S2 isInfluential citation
Referenced by (3)
- ATLAS cites — S2 isInfluential citation
- MemBART cites — S2 isInfluential citation
- Memformers (gradient memory) cites — S2 isInfluential citation