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.
Link graph surfaces - brainstorm audit trail
Produced by the v0.10.17 brainstorm pass on 2026-05-25. Primary
consultant: Claude Code CLI (docs/brainstorm/link-graph-surfaces/cli-output/claude.md).
Fallback consultant: not invoked (primary returned three parseable
variants + a clear recommendation on the first run).
Variant 1 - Single link-graph subsystem (strict precedent match)
Approach. Mint one new subsystem src/core/brain/link-graph/ that
hosts every pure helper for features 1-5 (alias resolver, anchor
extractor, unlinked-mention scanner, concept-cluster assembler, MOC
classifier). Atoms extend BacklinkRef (anchor + alias-source
fields) in place under backlinks.ts and add aliases to the
in-memory note record from vault.ts; consumers are new CLI verbs /
MCP tools that import from link-graph/. Feature 6 lands as
src/core/search/property-filter.ts inside the existing search
layer, and feature 7 lands as a small extension to whichever loader
already injects active.md into session context.
Trade-offs.
- Pro: Mirrors v0.10.15 / v0.10.16 exactly - one themed subdir, atoms/helpers/consumers DAG, predictable for reviewers.
- Pro: Helpers are co-located, so
synthesiseandmoc-auditcan re-use the alias / anchor / unlinked-mention helpers without cross-subsystem imports. - Pro: Search and session-context changes stay in their natural
layers instead of being dragged into a
brain/subdir where they don't belong. - Con:
link-graph/becomes a relatively large bag (5 helpers) compared to the tighter v0.10.16 trust/ tree. - Con: Two of seven features (property search, VAULT.md) sit outside the named release subsystem, so the release headline isn't quite "everything under one roof."
Complexity. medium Risk. low
Variant 2 - Three small co-equal subsystems (one per concern)
Approach. Carve the release into three named subsystems -
src/core/brain/graph/ for link / backlink helpers (1-5),
src/core/brain/property-filter/ for the FTS-post property-filter
phase (6), and src/core/brain/vault-context/ for the VAULT.md
loader (7). Each follows its own atoms/helpers/consumers DAG; the
CHANGELOG groups them under one v0.10.17 header but the code tree is
unambiguous about which file belongs to which feature cluster.
Trade-offs.
- Pro: Clean one-feature-cluster-per-subdir mapping; each subdir is small and individually testable.
- Pro: Property filter and VAULT.md are first-class members of the release, not "miscellaneous additions."
- Pro: Future bundles can grow any of the three without touching the others.
- Con: Breaks the "one new subsystem per release" precedent set by v0.10.15 and v0.10.16; three new subdirs in one PR is a noticeable convention drift.
- Con:
property-filter/underbrain/is awkward because the actual code is search-layer (FTS + chunks), not brain-layer; pulling it underbrain/muddies the layering. - Con: VAULT.md is a tiny addition that doesn't deserve its own subsystem; padding it out to atoms/helpers/consumers is ceremony for ~1 file of logic.
Complexity. medium Risk. medium
Variant 3 - In-place enrichment, single small subsystem only for read-views
Approach. Extend existing files directly for the structural
plumbing: wikilink.ts keeps anchors instead of stripping them,
backlinks.ts learns to resolve aliases when collecting refs,
search.ts grows a property-filter post-phase, and the existing
session-context loader learns about VAULT.md. Mint one small
subsystem src/core/brain/graph-views/ that only contains the
higher-level read-only assemblers (unlinked-mentions scanner,
concept-synthesis envelope builder, MOC audit classifier) - the
things that genuinely need shared helpers and dedicated tests.
Trade-offs.
- Pro: Smallest blast radius - file count is well under 70, structural changes live next to the code they enrich, no orphan one-file subsystems.
- Pro: The new subdir holds only the genuinely novel logic (the three "views"), so its scope is crisp and its test directory is well-bounded.
- Pro: Property filter living inside
src/core/search/matches the actual layer ownership, no awkward cross-tree imports. - Con: Diverges from the recent two-release precedent where ALL new helpers landed inside the new subdir; reviewers used to that pattern may need orientation.
- Con: Atom-level extensions are scattered across
wikilink.ts,backlinks.ts, andsearch.tsrather than visible in one diff stack, so the "what's the v0.10.17 shape change?" question takes more reading. - Con:
graph-views/doesn't capture the alias/anchor work in its name, so the release theme is split between "view helpers" and "graph plumbing upgrades."
Complexity. small Risk. low
Recommended: Variant 1
Consultant rationale. It is the only variant that holds the
line on the project's established two-release precedent (one named
subsystem with the atoms / helpers / consumers DAG), which keeps
reviewer mental model, test-directory layout, and CHANGELOG framing
identical to v0.10.15 / v0.10.16. The two features that genuinely
don't belong under brain/link-graph/ (property search is
search-layer, VAULT.md is session-context-layer) land in their
natural layers without inventing ceremonial subsystems for them,
avoiding Variant 2's drift while still keeping the headline
"link-graph subsystem" intact - and unlike Variant 3, the
foundational alias/anchor atoms remain visible as part of the new
subsystem's shape rather than scattered across pre-existing files.
Orchestrator decision
Accept the consultant's recommendation without override.
The variant aligns with three live constraints not yet captured in the prompt:
- The
pref-no-hardcoded-languagespreference applies to every helper in this bundle; co-locating the helpers under one named subsystem makes the language-agnostic review pass cheaper (one directory to audit, not three). - The CodeRabbit reviewer historically lands more findings when atoms drift across pre-existing files (Variant 3); concentrating them in the new subsystem reduces that surface.
- The release-notes excalidraw / PNG (precedent: trust/) renders as three layers (atoms / helpers / consumers). Variants 2 and 3 would require redrawing that layout to two or four bands. Variant 1 reuses the same three-layer template directly.
Property filter remains in src/core/search/property-filter.ts
(search-layer, natural ownership). The vault-root instruction file
lands in src/core/brain/vault-instruction-file.ts at brain-layer
root (it is consumed by brain_context, not by the link-graph
subsystem) - not under link-graph/ even though it ships in the
same release.