Skip to content

Memory

Status: mostly built

Built: the measured model profile (written by deputy bench, read back for the turn limit and by status), the trace of every run and feedback (reported by deputy stats), and lessons derived from the trace (deputy learn, see configuration). Lessons are tested against the benchmark (deputy lessons check). Project memory (pointers and a glossary, kept per repository) is built too, see configuration. The rest of this page is the design.

Does memory make sense? Yes, but narrowly. A model that is not trained on your project forgets everything between calls; the useful part of memory is what is measured and verified, not a diary. Three layers, each with a different owner and lifetime.

1. Model profiles (shared, machine-wide)

Measured facts about a model and its serving setup: tokens/s, the context that really works, tool-call reliability, best reasoning budget per task class, failure habits (for example: cites wider line ranges, asks for a shell). Written by the benchmark, never by a model. Highest value, lowest risk.

2. Lessons (shared and per project)

A lesson is a correction that was validated: "rule X added to profile Y raised line-citation accuracy from 81% to 96%". Fields: scope (model, task class, project), the rule text, evidence (run ids), confidence, expiry. Retrieved only when the task class matches, under a small token budget. Lessons that stop paying are removed.

3. Project memory (per repository)

Pointers first: where the rules live (.github/instructions/, AGENTS.md), the commands that verify, the glossary, past accepted diffs as examples. Point to the source of truth instead of copying it, so memory cannot go stale or contradict the repository. Stored in the user data directory keyed by repository id, exportable to reviewed markdown. Built: deputy memory.

Rules

  • Workers cannot write memory. Model profiles come from the benchmark, lessons from the trace (deputy learn), project memory from you (deputy memory); all are JSON files in the state directory, and project memory can be exported as markdown to review.
  • Lessons have evidence and an expiry, and their text comes from a fixed catalogue: nothing from a trace or a worker is copied into a prompt. No secrets, tokens or personal data.
  • A repository's memory is never added to another repository's task.
  • Ablation: a lesson stays only while a benchmark run with it is not worse than without it (deputy lessons check). Project memory has no ablation yet.
  • Selection is simple: lessons by profile and kind of check, project memory by repository, both under a small size budget. No search index; embeddings only if they ever win on the benchmark.