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
Memory Observability Suite - implementation plan
Tasks
Task 1: Versioned continuity schema (t_26040ee8)
- Files:
src/core/brain/continuity/types.ts(schemafield onContinuityRecord,CONTINUITY_SCHEMA_VERSIONconstant),src/core/brain/continuity/store.ts(stamp inbuildRecord),tests/core/brain/continuity-schema-version.test.ts - Acceptance: new records carry
schema: "o2b.continuity.v1"; readers accept legacy records without the field;recordId()provably excludes the schema field (same id for stamped and unstamped identical input) - Depends on: none
Task 2: Lazy gated emit kernel + call-site audit (t_5d7aa7c5)
- Files:
src/core/brain/continuity/emit.ts(new), call-site reroutes insrc/core/brain/context-pack.ts,src/core/brain/pre-compress-pack.ts,src/mcp/search-tools.ts(gate telemetry);tests/core/brain/continuity-emit.test.ts(helper semantics), no-consumer regression tests per gated surface (extendtests/core/brain/{context-pack,context-receipts}.test.ts,tests/mcp/recall-gate-telemetry.test.ts), plus a lifecycle boundary test (capture with defaults produces zero continuity writes) - Acceptance: with gate off, payload thunk never invoked and no continuity write happens (asserted per gated surface); a throwing thunk or write never fails the primary operation; gate on keeps existing record shapes byte-identical;
session-recall.ts/pre-compact-extract.tsdocumented as always-on primary writes and left fail-fast - Depends on: Task 1
Task 3: Continuity read-model
- Files:
src/core/brain/continuity/read-model.ts(new),tests/core/brain/continuity-read-model.test.ts - Acceptance: raw JSONL lines (stamped, legacy, malformed) normalize into one canonical shape; legacy records report schema v1; masking policy drops
privaterecords by default and masks per option; unknown kinds pass through fail-soft - Depends on: Task 1
Task 4: ATOF/ATIF mapping write-up + export verb (t_51959aeb)
- Files:
docs/brainstorm/memory-observability-suite/atof-atif-mapping.md(mapping, spec revisions, go decision),src/core/brain/continuity/export-atof.ts,src/core/brain/continuity/export-atif.ts,src/cli/brain/verbs/continuity.ts, verb registration (verbs/index.ts,brain.ts,command-manifest.ts,help-text.ts),tests/core/brain/continuity-export-atof.test.ts,continuity-export-atif.test.ts,tests/cli/continuity-export.test.ts(golden files with sanitized fixtures) - Acceptance:
o2b brain continuity export --format atof|atif [--session <id>] [--month <m>] --out <dir>produces spec-conformant files;privaterecords dropped,redactedtext stays masked; golden-file tests pass; read-only over the store - Depends on: Task 3
Task 5: Bench harness core (t_882c396a, part 1)
- Files:
src/core/bench/types.ts,src/core/bench/fixture.ts,src/core/bench/run-store.ts,src/core/bench/phases.ts,src/core/bench/report.ts,src/core/bench/judge.ts,tests/core/bench/{fixture,run-store,phases,report}.test.ts - Acceptance: fixture JSON validates and materializes a disposable vault inside the run directory; phases ingest/index/retrieve/evaluate/report run end-to-end deterministically with no network; checkpoint/resume by run ID skips completed phases and invalidates on fixture-hash change; report keeps quality/latency/context-cost separate with stable sorted keys; guard refuses paths outside the run dir
- Depends on: Task 3 (session-handoff fixture reads normalized continuity records)
Task 6: Bench CLI verb + fixture suite (t_882c396a, part 2)
- Files:
src/cli/brain/verbs/bench.ts, verb registration,tests/fixtures/bench/*.json(6 categories),src/core/config.ts(resolveBenchJudgeCmd+ env override),tests/helpers/run-cli.ts(env scrub),.gitignore,tests/cli/bench.test.ts - Acceptance:
o2b brain bench memory --fixture <name> [--resume <run-id>] [--json]runs the full pipeline on every shipped fixture; at least one fixture catches stale/superseded recall regressions; at least one measures expected evidence within context budget; judge phase skipped cleanly whenbench_judge_cmdunset - Depends on: Task 5
Task 7: e2e integration test
- Files:
tests/e2e/memory-observability.integration.test.ts - Acceptance: one flow exercises schema-stamped emission with gates on, export of the produced records to both formats, and a bench run over a small fixture - all green in one test file
- Depends on: Tasks 2, 4, 6
Task 8: Observability contract doc (t_66545537)
- Files:
docs/observability.md, cross-links fromdocs/how-it-works.mdanddocs/mcp.md - Acceptance: every emitted event kind enumerated and verified against source (log events, continuity kinds, session lifecycle); opt-in vs always-on correct per kind; correlation IDs, payload safety, fail-open rules, schema version + evolution rule documented; repo writing prefs honored
- Depends on: Tasks 1, 2 (documents their behavior)