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.
You are brainstorming architectural variants for the following task. Do not write code. Do not write a final design. Only produce variants and a recommendation.
Task
Two kanban cards ship together in one release wave, "Dated commitments, capture to surface":
Card t_da8673aa (validator-rescoped 2026-09-12): automatic open-loop / commitment detection with due dates and evidence quotes. Rescoped scope: a due-date and evidence-quote vocabulary on a one-off commitment page family; an agent-supplied extraction intake mirroring the existing brain_agenda and brain_intake_entities tool contracts (the kernel never calls a model); projection into facts and entity edges. Gmail/Calendar/Contacts ingestion is out of scope (vault-notes model). Today, open loops are manual @osb loop markers with no due date and no store, and the only due-dated pages are cadence-driven recurring obligations, so a one-off promise in prose is invisible.
Card t_43b0a379 (validator-rescoped): opt-in counterparty brief injecting dated known-person facts at session start and on reply or meeting intent. Rescoped scope: a push-side Before-you-start hook block with at most three dated, shareable items for a confidently matched person or company (off by default, at most once per session, degrades to a pull-style context pack call on harnesses without hooks), plus a pull-side entity_facts section of context packs (recent facts per packed entity, no dependency on the dated-item supply). Privacy posture: only shareable items may be briefed; private pages excluded. The dated-item supply depends on card t_da8673aa landing first.
Project context
Open Second Brain: a memory layer for AI agents living in an Obsidian vault; TypeScript on Bun; Markdown pages under Brain/ are the source of truth; MCP tools, lifecycle hooks and an o2b CLI are the surfaces. Deterministic kernel: the kernel never calls a model.
Recent commits: f203591e feat: measured autonomy - a rule-promoting dream pass, model-aware context budgets and honest benchmarks (#269) 9e579820 feat: note-change truth maintenance for links, index freshness and entity anchoring (#268) f7c1c88c feat: entity merge repair, write destinations, cross-agent catch-up and OKF v0.2 (v1.88.0) (#267) f6d9f95f feat: guarded host writes, walkable list results, a converse decision route and a snoozable trigger queue (v1.87.0) (#266)
Related files (all anchors verified against the live tree this session):
- src/core/brain/open-loops.ts -
@osb loopmarkers: OpenLoop is { id, text, path, line } only; open set re-derived per call, no store; strictly deterministic and frozen. - src/core/brain/obligations.ts - recurring obligations, the due-dated page-family precedent: pages at Brain/obligations/.md, pure UTC cadence math, add/done/list/show/remove, list due-sorted with overdue flag and days-until-due, remove archives the page.
- src/core/brain/preference.ts:213-218 and src/core/brain/commitment.ts - NAMING COLLISION:
commitmentis an existing user-editable frontmatter tier (exploring | leaning | decided | locked) on preferences, theses and decisions, with its own validator and tolerant reader. A new page family about promised work needs a distinct family name and frontmatter keys that never reuse the barecommitment:key. - src/mcp/brain/calendar-tools.ts - brain_obligation (one tool, operation enum add/done/list/show/remove) and brain_agenda (stateless analysis over caller-supplied events; pure, no vault writes).
- src/core/brain/intake/extract-intake.ts and src/mcp/brain/ner-tools.ts - the agent-executed extraction-intake precedent: the agent owns the model and posts a typed payload; the kernel validates the whole payload and refuses BEFORE writing anything; commits idempotently through duplicate-refusing upserts; stamps provenance; trust is derived from cited sources, never an argument.
- src/core/graph/relation-vocab.ts - DEFAULT_RELATION_TYPES (related, extends, depends_on, refines, contradicts, superseded_by); data-driven: adding a relation type is a one-line change, and a frontmatter field whose name is a known relation becomes a typed edge automatically (src/core/graph/frontmatter-relations.ts, ingested by the search indexer and the link-graph).
- src/core/brain/atomic-facts.ts:157 - anchorEntityForms: the deterministic, quality-gated anchoring kernel matching entity name and alias forms against prose (minimum form length, clean leading edge, separator-tolerant variants); canonical scope filtering via entityStatusInScope.
- src/core/graph/visibility.ts - the one reserved visibility token
privatedominates any co-tag; strictestVisibilityOf computes the intersection for items derived from several source pages. - src/core/brain/context-pack.ts - the tier-then-recency context pack; conditional-option precedent:
densityRanking?: booleanoption (config key density_ranking_context_pack, default off, resolved at the tool seam), a conditional report member, and a conditional output projection, so default packs stay byte-identical; a synthesized attention block shows how a non-page item joins the pack; four callers share packContext (MCP tool, anticipatory cache, retrieval plan, bench). - hooks/hooks.json, hooks/gap-agenda.ts, hooks/recall-inject.ts - SessionStart runs five hooks, UserPromptSubmit runs four; a new hook is one hooks/.ts plus a hooks.json row; opt-in flags are machine-local config booleans with env overrides (resolveConfigFlag), checked FIRST so flag-off is a byte-identical no-op; the injection ledger gives once-per-session suppression; only SessionStart and UserPromptSubmit may emit additionalContext; hooks run under a hard process ceiling.
- src/core/config.ts:852 - resolveDensityRankingContextPack, the config-key-at-tool-seam precedent.
Conventions:
- The kernel never calls a model; the calling agent supplies extractions, the kernel validates typed payloads.
- No hardcoded natural-language word lists anywhere: operator-declared vocabularies or structural markers only; all strings are English and other-language cases are handled abstractly.
- Errors surface explicitly by name; no silent fallbacks, no stubs; repeated values hoisted into named constants; withheld or unreadable pages answer as absent (identical-to-absent), never counted.
- New MCP tools are deliberate surface changes: frozen tool-name parity test, a visibility-surface-registry row plus a two-directional census, reach predicates on every read, and vault-path-safe responses. A new hook must register in hooks.json plus three test-enforced mirrors (codex mirror, grok parity, shape test).
- Each new page family follows the obligations pattern: path constants, path helpers, a core module with a typed error, write-time validators, a soft reader for hand-edited frontmatter, deterministic render/parse, injected clock.
Constraints:
- One PR, roughly 50-70 changed files including tests and docs; max features = the two cards.
- The extraction contract is caller-executed: nothing may re-derive commitments by walking notes with a model, and nothing may require a model at write time.
- Alias and intent matching must stay deterministic and re-derived server-side.
- Intent (reply/meeting) detection must not be a hardcoded natural-language word list: an operator-declared vocabulary in config or an explicit structural marker are the discipline-compatible options.
- The push side must not start before the dated-item supply exists (t_da8673aa unblocks t_43b0a379).
- The pull side is dependency-free and should be implementable as an option on brain_context_pack rather than a new tool (avoiding a four-census tool registration).
Required output format
Produce exactly 3 distinct architectural variants. For each variant:
Variant N:
- Approach: 2-3 sentences describing the variant.
- Trade-offs: bullet list of pros and cons.
- Complexity: small | medium | large
- Risk: low | medium | high
After the three variants, add exactly one recommendation:
Recommended: Variant N
Rationale: 2-3 sentences explaining why this variant over the others, considering the project context and constraints above.
Output nothing outside of these sections.