RazorAttention
https://openreview.net/forum?id=tkiZQlL04w
Efficient KV-cache compression through retrieval heads.
At a glance
- Type
- KV-cache compression via retrieval heads
- Tier
- T3
- Created
- 2024-07
- 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
- kv-cache
- retrieval
- attention
- persistence
- session
- update
- evict-oldest
- unit
- kv-token
- governance
- n/a
- conflict
- n/a
When to use
Optimised for: not applicable - research paper
Anti-fit: not applicable - research paper
Pros & cons
Pros
Aggressive KV pruning for long context with retention of high-importance tokens.
Cons
Research-stage; pruning quality bounded by importance heuristic.
Claims & capabilities
Training-free KV cache compression that preserves all token information using head-specialized strategies — full caches for retrieval heads, discarding remote tokens in non-retrieval heads, with compensation tokens; over 70% reduction in KV cache size; 3X reduction under general cases; no noticeable performance impact; FlashAttention-compatible
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 recent method papers — theorized, no distinct product in the catalog, ranked by inbound references.
- Compressive Transformer T3
Maintains recent states in full resolution while compressing older memories with learned compression functions. DeepMind.
- MemGPT v2 / agent-tools T3
Already in catalog as the foundational MemGPT paper. Note: Letta is the productionised successor (cross-listed).
- Transformer-XL T3
Extends context through segment-level recurrence + caching of hidden states from prior segments. Foundational long-context architecture.
- Generative Agents T3
Park et al. — landmark agent-simulation paper. Reflection + memory stream + retrieval enable believable agent behavior.
- MemoryBank T3
Enhances LLMs with long-term memory. Early influential paper.
- Reflexion T3
Language agents with verbal reinforcement learning.