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
4.7 KB
Raw Blame

Session Knowledge Synthesis - variant audit trail

How scope was chosen

This cycle the operator asked for a large scope of related, low-priority triage tasks, explicitly excluding "adding new agents" (neither install adapters like cline nor session-import adapters like copilot-cli / pi), and asked to move already-implemented-but-open tasks to done.

Phase 0 ran the CLI consultant (claude -p) over the entire triage column (cli-output/prompt.md, cli-output/claude.md). The consultant recommended the Session Knowledge Synthesis cluster (t_325a7e4a, t_e4ddbe7c, t_635a3ea5) with Variant 1 (agent-extracted structured records + read-side synthesis tools, kernel stays provider-agnostic). Full output is preserved verbatim in cli-output/claude.md.

Orchestrator gap analysis (override of the consultant cluster)

The consultant worked from the triage text and a context summary; it did not know the codebase had moved ahead of two of the three tasks. Direct inspection of main found:

  • t_e4ddbe7c (daily / morning brief) is already shipped. brain_brief exposes view=morning (src/core/brain/morning-brief.ts:buildMorningBrief) and view=daily (src/core/brain/temporal/daily-brief.ts:buildDailyBrief), plus weekly / monthly, from the temporal-synthesis suite. The only novel upstream element - injecting the brief into a task tracker's description field - is host-specific and not an Open Second Brain kernel concern. -> Closed as already-present (moved to done), no code.
  • t_325a7e4a (structured session summary) is only partially covered. pre_compact_extract already does agent-supplied structured extraction (decision / commitment / outcome / rule / open_question), and session_summary_node is a hierarchical recall rollup. Neither is a session-scoped digest over the four canonical categories (request / decisions / learnings / next_steps). -> Real but narrow gap; build it as a thin session-scoped envelope reusing the existing extraction + storage primitives.
  • t_635a3ea5 (idea lineage) is a real gap. The edges exist (sourceRefs, belief-evolution, dream output) but no read-side tracer reconstructs the observation -> synthesis -> conclusion chain. -> Build as a read-only tracer.

To keep a meaningful scope after removing the already-done task, the operator agreed to add a third in-repo, non-agent task:

  • t_6a201155 (episodic note-history decomposition). Substrate confirmed: src/core/brain/git/reader.ts (Project History Suite) is a sanitized read-only git reader, shipped and tested but with no current callers. A note's version chain = the git commits touching its path; phases are a deterministic structural split. -> Build as its first consumer.

Final build scope: A structured session summary, B idea lineage, C episodic note history. t_e4ddbe7c closed as already-present.

Architectural variants considered (from the consultant, applied to the final scope)

Variant 1: Agent-extracted structured records + read-side synthesis tools

  • Approach: new structured record(s) written by an MCP tool the agent calls; read-side tools aggregate. Kernel stores already-extracted data, never calls an LLM.
  • Trade-offs: matches the provider-agnostic guarantee exactly; maximally additive and byte-identical when unused; language-agnostic by construction. Con: rich fields only populate when an agent is active.
  • Complexity: medium. Risk: low.

Variant 2: Deterministic skeleton + optional agent enrichment (sentinel-marked)

  • Approach: kernel writes a deterministic structural skeleton with no provider; AI fields fill later and are marked derived.
  • Trade-offs: strongest fail-soft story. Con: two-phase lifecycle, more ordering to reason about; couples to the dream pass timing.
  • Complexity: medium. Risk: low.

Variant 3: Unified TemporalSynthesisEnvelope (refactor weekly-brief)

  • Approach: generalize WeeklySynthesisEnvelope into one envelope parameterized by scope.
  • Trade-offs: most DRY long-term. Con: refactors a shipped surface, threatening byte-identical-when-unused; largest blast radius.
  • Complexity: large. Risk: medium.

Final decision

Variant 1, borrowing Variant 2's discipline: every AI-derived field is optional and marked derived (enriched flag / absent digest), so no-provider cases degrade to a deterministic skeleton or honest absence, never a fabricated summary. Variant 3's unification is rejected - refactoring the shipped weekly-brief surface is unjustified risk for an additive release. This applies uniformly to all three features: the kernel validates and stores agent-supplied structure (A), reads existing edges (B), and reads git history with a deterministic split plus optional enrichment (C).