Anthropic Circuit Tracing
https://transformer-circuits.pub/2025/attribution-graphs/methods.html
Replaces MLP layers with sparse cross-layer transcoders to produce "replacement models" where activations correspond to interpretable features; traces attribution graphs showing which features causally influenced which outputs. Includes case studies on factual recall, revealing how facts are stored and retrieved from weights at the circuit level. Open-source tools run on any open-weights model.
At a glance
- Type
- Mechanistic interpretability of parametric storage
- Tier
- T2
- Created
- 2025-03-27 (published March 27 2025 on transformer-circuits.pub)
- 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
- parametric-recall
- persistence
- parametric-permanent
- update
- read-only
- unit
- weight
- governance
- auditable
- conflict
- n/a
When to use
Optimised for: not applicable - research paper
Anti-fit: not applicable - research paper
Pros & cons
Pros
Only published tool that makes parametric (weight-level) memory storage mechanistically interpretable rather than behaviorally probed; attribution graphs are auditable per-prompt.
Cons
Not a memory system — it studies memory; findings are descriptive, not prescriptive; scalability to full production Claude models is limited (primary validation on 18-layer models and Haiku).
Claims & capabilities
Successfully traced computational circuits for factual recall, planning, and multi-step reasoning in Claude 3.5 Haiku; open-source library released.
Technical surface
- API surface
- searched not found
- Backend storage
- searched not found
- Deployment
- not applicable — not a deployable product
- Embedding model
- searched not found
- Multi-tenancy
- searched not found
- 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.