Whitepaper

The AI Memory Prosthesis (SR-SI)

A technical methodology for making AI agents reconstruct project context before they act.

LLMs are fluent, capable, and structurally forgetful. They can sound coherent in a long session while the project map underneath them has already started to drift.

SR-SI, Simulated Recall via Shallow Indexing, treats that as an architecture problem. The agent does not need to remember the whole project inside the conversation. It needs a compact, maintained index that tells it where to look, what matters, and which constraints shape the current task.

The paper documents the method behind that operating memory: shallow indexes, scoped source loading, index-first execution, and agent-operated maintenance under human policy control.

What the paper argues

Context decay is not only a model limitation. It is the result of asking a stateless conversational system to carry architectural memory through accumulated dialogue.

SR-SI changes the collaboration model. The record becomes part of the workflow, not a document someone might read later. The agent consults the index before work, acts inside the current system, and updates the record when the system changes.

What changed in v2

85.5%

Reduction in net context-maintenance overhead in the measured SR-SI retrofit project.

7x

Token-efficiency improvement over the unstructured conversational baseline in the original paper.

9.5x

Additional improvement after splitting one shallow index into scoped sub-indices on a larger production codebase.

106x

Approximate cumulative token-coherence improvement against the unstructured baseline, using the paper's TCM framing.

These figures are applied production evidence from Moe's own codebases. They are not presented as an industry benchmark, and independent validation remains the next useful test.

How SR-SI works

The mechanism is intentionally simple. The strength comes from enforcing the loop, keeping the index shallow, and making consultation mandatory before action.

  1. 01 Read the master index before acting.
  2. 02 Load only the scoped source of truth needed for the task.
  3. 03 Execute inside the current contracts.
  4. 04 Update the index when behavior or ownership changes.
  5. 05 Keep human control at the policy layer.

Evidence and limits

The original work showed the practical value of moving from conversation memory to reconstructable context. The later modular version showed that the same idea scales better when the master map stays lean and scoped sub-indices carry domain detail.

The limit is also clear. SR-SI does not give an LLM inner memory, subjective continuity, or permanent knowledge. It creates procedural continuity: a reliable way for the system to rebuild the right working context from artifacts each time.

Related reading

SR-SI: the methodology that gives AI persistent memory across any long-running project What SR-SI actually is, and why I stopped explaining it as an AI trick Context decay in AI is not a model problem Memory is retrieval strategy, not storage

Next step

Stop guessing. Move to execution.

Build durable AI workflows