Skip to content

Local-first 🧠 memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.

TypeScript
3.6 KB
Raw Blame

Recall & Working-Memory Quality Suite - brainstorm audit trail

Primary consultant: Claude Code (claude -p), exit 0, three parseable variants plus a recommendation. Fallback consultant (Codex) NOT invoked - primary succeeded with >= 2 variants. Full verbatim output: cli-output/claude.md.

Variant 1: Federated patches with a shared config block (small / low)

Each unit a self-contained module; only shared structure is one optional nested block on ResolvedSearchConfig plus resolveRecallProfile. Read-side weighting for decay and file-context written inline at the score-combination site in search.ts.

  • Pro: smallest diff; independently testable/revertable; trivial byte-identical OFF; no new abstraction to get wrong.
  • Con: read-side logic for Units 3 and 4 scattered and order-sensitive, no single seam; persistence conventions re-implemented per unit; reads as four patches that merely landed together.

Variant 2: Two shared primitives - profile/knob resolver + composable read-side adjuster fed by signal readers (medium / medium) - RECOMMENDED

A resolveRecallProfile reusing tuning.ts axes, plus a read-side ScoreAdjuster composition point fed by small UsageSignalReaders over recall_telemetry. Co-occurrence stays a dream-side producer sharing only the persistence convention.

  • Pro: genuine composition on the one axis that overlaps (read-side weighting); shared signal readers + persistence convention; clean identity-transform OFF.
  • Con: needs up-front adjuster-contract/ordering design; two seams; a middle ground.

Variant 3: Unified RecallContext orchestrator with internal signal/pass registry (large / high)

A single buildRecallContext owns profile expansion, signal registration, adjuster composition, file-context path, and a maintenance-pass registry for co-occurrence.

  • Pro: maximal uniformity; most natural growth toward future recall sources.
  • Con: heavy abstraction for four small units; sits on the hot search() public path (enlarges blast radius, harder byte-identical proof); drifts toward the explicitly-deferred external recall-source registry; highest review cost.

Consultant recommendation: Variant 2

It matches the stated goal of one or two shared primitives precisely ... the units compose on the single axis they truly share (read-side weighting) while co-occurrence stays a dream-side producer behind only the shared persistence convention ... deliberately stops short of the orchestrator/registry that would drift into the explicitly deferred external-source scope.

Orchestrator decision: ADOPT Variant 2, with one scoping refinement

Accepted. Variant 2 fits the project's established conventions (flag-gated, deterministic, versioned+hashed persistence, byte-identical OFF) and the "one or two shared primitives" goal better than the minimal Variant 1 (which re-implements persistence per unit and scatters read-side logic) or the heavy Variant 3 (which puts an orchestrator on the hot search path and drifts toward the deferred external-source registry).

Refinement (documented in design.md): the consultant placed the read-side adjuster for BOTH decay and file-context on the search() public path. We keep the two-primitive philosophy but scope the read-side weight to where candidates actually flow - decay weighting in the continuity read-model, file-context as a standalone surface reusing session-focus.ts. This leaves the search() default path untouched, making the byte-identical guarantee cheaper and the blast radius smaller - which addresses the consultant's own stated risk about widening the search hot path. This is a scoping choice within Variant 2, not a different variant.