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.
Part 1 - Scope recommendation
Candidate clusters
- Session Knowledge Synthesis —
t_325a7e4a(p0, structured session summary),t_e4ddbe7c(p3, daily morning brief),t_635a3ea5(p2, idea-lineage). Theme:src/core/brain/temporal/+session-lifecycle.ts— turn session activity into structured, queryable, surface-able temporal knowledge via agent-driven extraction. Usefulness: high (closes the "what happened / what was decided / what's next" loop, the core value of a second brain); risk: low, fully additive. - Multi-agent session-import adapters —
t_aaced180(p1, Copilot history),t_f5761ca8(p1, Pi),t_7cf20134(p1, Copilot CLI),t_7c89c306(p1, progressive-batch parser). Theme:src/core/brain/sessions/adapter registry. Usefulness: moderate; risk: blocking — every adapter requires reverse-engineering an external/unstable transcript format (Pi explicitly "needs reverse-engineering"; Copilot format external), which the brief flags as a blocking risk. - Graphify/codegraph passthroughs —
t_b8283326,t_bf6933bc,t_5d8d6a00,t_85252236,t_a286135c,t_6c9f4434,t_a1e76788. Theme:src/core/partner/codegraph.ts. Usefulness: low-moderate; risk: blocking — these are thin passthroughs whose logic lives in the externalgraphifybinary and would silently no-op when it is absent (t_bf6933bcwas already de-scoped for exactly this misleading-fallback reason). - MCP server distribution / transport —
t_da6321a9(p3, console-script entry),t_31dfae18(p2, Streamable HTTP transport). Theme:src/mcp/. Usefulness: low for a memory store; risk: medium — validator notes the two are "different layers," and HTTP transport + API-key auth leans architectural (multi-client security model).
Recommended cluster
- Name: Session Knowledge Synthesis
- Task IDs:
t_325a7e4a,t_e4ddbe7c,t_635a3ea5 - Why this cluster (usefulness-first): For a second brain, the highest-leverage gap is that sessions are stored but not queryable as decisions/learnings/next-steps —
t_325a7e4a(the only cleanly-buildable p0) fixes exactly that, producing machine-queryable structured memories instead of opaque transcripts.t_e4ddbe7cis the natural read-side consumer: a daily brief that surfaces those structured summaries plus pending items, completing a coherent session→summary→brief loop.t_635a3ea5adds the provenance leg, tracing how a decision or learning was reached, which makes the structured summaries auditable and trustworthy. All three live insrc/core/brain/temporal/andsession-lifecycle.ts, all rely on agent-driven extraction with the kernel storing already-structured data, and all are byte-identical-when-unused and language-agnostic. This is far more useful to the product than adding one more editor or codegraph passthrough. - Why not the others:
- Multi-agent session-import adapters: blocked — each depends on reverse-engineering an external/unstable transcript format.
- Graphify/codegraph passthroughs: blocked — thin passthroughs to an external binary that would silently no-op when absent (forbidden misleading fallback).
- MCP distribution/transport: loosely related ("different layers"), HTTP transport leans architectural, and it adds little to recall quality.
- Key risks and how to neutralize them:
- Provider dependency for AI extraction. Neutralize by keeping all AI-derived fields optional and clearly provenance-marked; absence of a provider degrades to a deterministic structural skeleton, never a fake summary.
- Touching shipped
weekly-brief.tsbehavior. Neutralize by adding the daily variant additively (new cadence/scope), guaranteeing byte-identical output when the daily surface is unused. - Language-agnostic guarantee. Neutralize by extracting categories (decisions/learnings/next-steps) via the agent/LLM and structural signals only — no hardcoded natural-language keyword lists.
- Scope creep into provenance graph. Neutralize by scoping idea-lineage to recording edges over existing belief-evolution/dream output, not a new graph engine.
Part 2 - Architectural variants for the recommended cluster
Variant 1: Agent-extracted structured records + read-side synthesis tools
- Approach: Add a
session_summarymemory type with structured metadata (request/decisions/learnings/next-steps), written by an MCP tool the agent calls at session end (or via the existing stop hook prompting the agent). The daily brief and idea-lineage are read-side synthesis tools that aggregate these records; the kernel only stores already-extracted structured data and never calls an LLM. - Trade-offs:
- Pro: Matches the provider-agnostic kernel guarantee exactly; AI extraction stays fully agent-driven.
- Pro: Maximally additive — new memory type + new tools, byte-identical when unused; minimal contact with shipped
weekly-brief.ts. - Pro: Language-agnostic by construction (agent does extraction, no word lists).
- Con: Rich fields only populate when an agent/provider is active; raw-only sessions yield thin summaries.
- Con: Three loosely-shared tools rather than one unified pipeline (mild duplication).
- Complexity: medium
- Risk: low
Variant 2: Deterministic skeleton + optional agent enrichment (sentinel-marked)
- Approach: The stop hook deterministically writes a structural skeleton (turn count, files touched, tool calls) for every session with no provider needed; AI fields (decisions/learnings/next-steps) are filled later by an agent pass (e.g., during dream) and marked as derived using a generated/user sentinel pattern. Daily brief and lineage read from these progressively-enriched records.
- Trade-offs:
- Pro: Strongest fail-soft story — useful output even with no provider, with AI fields honestly marked derived (no misleading fallback).
- Pro: Sentinel marking enables safe re-synthesis without clobbering enriched content.
- Con: Two-phase lifecycle (skeleton then enrichment) is more moving parts and ordering to reason about.
- Con: Couples the feature to the dream/synthesis pass timing.
- Complexity: medium
- Risk: low
Variant 3: Unified TemporalSynthesisEnvelope (refactor weekly-brief into a general pipeline)
- Approach: Generalize the existing
WeeklySynthesisEnvelopeinto oneTemporalSynthesisEnvelopeparameterized by scope (session | daily | weekly) and synthesis type, so session summaries, the daily brief, and lineage are all variants of a single schema and pipeline. - Trade-offs:
- Pro: Most DRY; one synthesis path to maintain and extend later.
- Pro: Consistent schema across all temporal outputs.
- Con: Refactors shipped
weekly-brief.ts, directly threatening the byte-identical-when-unused guarantee for an existing surface. - Con: Largest blast radius; couples three new features to a refactor that must preserve v1.x weekly behavior exactly.
- Complexity: large
- Risk: medium
Recommended: Variant 1
Rationale: Variant 1 aligns cleanest with the project's hard guarantees — the kernel never calls an LLM, extraction is agent-driven, and a new memory type plus read-side tools are byte-identical when unused with almost no contact with shipped weekly-brief.ts. It should borrow Variant 2's discipline of marking AI fields as optional/derived so absent-provider cases degrade honestly rather than fabricating summaries. Variant 3's unification is appealing long-term but its refactor of a shipped surface is unjustified risk for an additive low-priority release.