Skip to content

docs: update v0.1.142 g5e handoff #307

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
cafitac merged 1 commit into from
May 12, 2026
Merged

docs: update v0.1.142 g5e handoff #307

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 19 additions & 19 deletions .dev/roadmap/memory-consolidation/current-progress-and-next-steps.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,20 +1,20 @@
# Memory Consolidation Current Progress and Next Steps

Status: AI-authored draft. Not yet human-approved.
Last updated: 2026-05-13 00:07 KST
Last updated: 2026-05-13 00:53 KST

## v0.1.141 + G5d completed checkpoint and next five-step runway
## v0.1.142 + G5e completed checkpoint and next five-step runway

This document is the restartable checkpoint after the v0.1.141 release/runtime rollout, fresh G4 diagnostics, merged G5a/G5b/G5c reviewed-candidate/scoring runway, and completed G5d read-only repeated activation -> reinforcement refinement preview.
This document is the restartable checkpoint after the v0.1.142 release/runtime rollout, fresh G4 diagnostics, merged G5a/G5b/G5c/G5d reviewed-candidate/scoring/reinforcement runway, and completed G5e read-only stale weak evidence -> decay/collapse candidate preview.

Current verified release state:

- Release: `v0.1.141`.
- GitHub Release: `https://github.com/cafitac/agent-memory/releases/tag/v0.1.141`.
- npm: `@cafitac/agent-memory@0.1.141`.
- PyPI: `cafitac-agent-memory==0.1.141`.
- Runtime: `/Users/reddit/.agent-memory/runtime/v0.1.141/.venv/bin/agent-memory`.
- Runtime smoke report: `/Users/reddit/.agent-memory/runtime/v0.1.141/g5d-live-smoke.json`.
- Release: `v0.1.142`.
- GitHub Release: `https://github.com/cafitac/agent-memory/releases/tag/v0.1.142`.
- npm: `@cafitac/agent-memory@0.1.142`.
- PyPI: `cafitac-agent-memory==0.1.142`.
- Runtime: `/Users/reddit/.agent-memory/runtime/v0.1.142/.venv/bin/agent-memory`.
- Runtime smoke report: `/Users/reddit/.agent-memory/runtime/v0.1.142/g5e-live-smoke.json`.
- Fresh G4 report directory retained: `/Users/reddit/.agent-memory/reports/g4-v0138-20260512-132253/`.

Fresh diagnostics:
Expand All @@ -23,25 +23,25 @@ Fresh diagnostics:
- `fresh-epoch-v0138.json`: `quality_gate.pass=true`, decision `fresh_epoch_ready_to_compare_against_historical`.
- `g4-review-queue-preview-v0138-fresh.json`: `quality_gate.pass=true`, decision `review_queue_ready_for_manual_review`, `read_only=true`, `mutated=false`.
- `scheduled-dry-run.json`: historical/full-window broad G4 still blocks on `trace_quality_needs_more_dogfooding`, `decay_risk_above_threshold`, and `background_quality_warnings_present`.
- G5a/G5b/G5c/G5d: `dogfood trace-cluster-preview`, `dogfood trace-candidate-persist/list/update/apply`, read-only `review_score`/`review_recommendation`, and `dogfood reinforcement-refinement-preview` are merged/released through v0.1.141.
- G5d reinforcement-refinement preview does not persist review state, increment reinforcement counts, promote memories, auto-approve ordinary conversation, or change retrieval defaults. Full release CI, publish, manual true-distribution smoke, and live Hermes runtime rollout passed.
- G5a/G5b/G5c/G5d/G5e: `dogfood trace-cluster-preview`, `dogfood trace-candidate-persist/list/update/apply`, read-only `review_score`/`review_recommendation`, `dogfood reinforcement-refinement-preview`, and `dogfood decay-collapse-preview` are merged/released through v0.1.142.
- G5e decay-collapse preview does not persist review state, delete/deprecate/collapse memories, promote memories, auto-approve ordinary conversation, or change retrieval defaults. Full release CI, publish, manual true-distribution smoke, and live Hermes runtime rollout passed.

Progress estimate:

- Overall north-star: 60-62%.
- Substrate/evidence plumbing: about 74-76%.
- Safe automatic mutation/promotion: about 42-45%.
- Remaining work: about 38-40% overall.
- Overall north-star: 62-64%.
- Substrate/evidence plumbing: about 75-77%.
- Safe automatic mutation/promotion: about 43-46%.
- Remaining work: about 36-38% overall.

Current interpretation:

Fresh v0.1.141 evidence and merged G5a/G5b/G5c/G5d are healthy enough to continue the brain-like reviewed-candidate runway. Broad G4/background apply remains blocked. G5d added repeated activation -> reinforcement refinement candidates as a ref-safe preview only; it is not approval for persistence, promotion, auto-approval, reinforcement mutation, decay, supersession, or retrieval-default changes.
Fresh v0.1.142 evidence and merged G5a/G5b/G5c/G5d/G5e are healthy enough to continue the brain-like reviewed-candidate runway. Broad G4/background apply remains blocked. G5e added stale weak evidence -> decay/collapse candidates as a ref-safe preview only; it is not approval for persistence, promotion, auto-approval, reinforcement mutation, decay/collapse mutation, supersession, or retrieval-default changes.

Recommended sequence from here:

1. Start G5e: stale weak evidence -> decay/collapse candidate preview, still preview/review-first and read-only.
2. Preserve G5e safety shape: review scores and recommendations remain review-priority signals only; they must not delete, decay, collapse, persist review state, promote memories, auto-approve ordinary conversation, or change retrieval defaults.
3. If a later G5d/G5e slice introduces mutation, keep it behind a separate explicit apply policy with backup, audit, approval phrase, actor, reason hash, and rollback.
1. Start conflict -> supersession/replacement candidate preview, still preview/review-first and read-only.
2. Preserve G5d/G5e safety shape: review scores/recommendations, reinforcement refinement, and decay/collapse candidates remain review-priority signals only; they must not delete, decay, collapse, persist review state, promote memories, auto-approve ordinary conversation, or change retrieval defaults.
3. If a later conflict/supersession or decay/collapse slice introduces mutation, keep it behind a separate explicit apply policy with backup, audit, approval phrase, actor, reason hash, and rollback.
4. G4 broad apply contract: preserve explicit policy/approval/actor/reason/backup/expected-queue/audit/rollback requirements and keep raw-content/default-retrieval/ordinary-auto-approval forbidden.
5. Historical telemetry reconciliation: use only a reviewed telemetry-only `telemetry-reset-v1` corridor for historical rows older than a selected epoch; protected memory tables must not mutate.

Expand Down
60 changes: 30 additions & 30 deletions .dev/status/current-handoff.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,43 +1,43 @@
# agent-memory current handoff

Status: AI-authored draft. Not yet human-approved.
Last updated: 2026-05-13 00:07 KST
Last updated: 2026-05-13 00:53 KST

## v0.1.141 + G5d completed checkpoint
## v0.1.142 + G5e completed checkpoint

Use `.dev/status/next-agent-memory-action.md` as the shortest current source of truth.

Current verified state:

- Latest completed release/runtime rollout: `v0.1.141`.
- Runtime: `/Users/reddit/.agent-memory/runtime/v0.1.141/.venv/bin/agent-memory`.
- Runtime smoke report: `/Users/reddit/.agent-memory/runtime/v0.1.141/g5d-live-smoke.json`.
- GitHub Release: `https://github.com/cafitac/agent-memory/releases/tag/v0.1.141`.
- npm/PyPI latest verified as `0.1.141`.
- Hermes configs updated from v0.1.140 to v0.1.141 and backed up as `/Users/reddit/.hermes/config.yaml.bak-v0141-20260513T000411` plus matching profile backups.
- Latest completed release/runtime rollout: `v0.1.142`.
- Runtime: `/Users/reddit/.agent-memory/runtime/v0.1.142/.venv/bin/agent-memory`.
- Runtime smoke report: `/Users/reddit/.agent-memory/runtime/v0.1.142/g5e-live-smoke.json`.
- GitHub Release: `https://github.com/cafitac/agent-memory/releases/tag/v0.1.142`.
- npm/PyPI latest verified as `0.1.142`.
- Hermes configs updated from v0.1.141 to v0.1.142 and backed up as `/Users/reddit/.hermes/config.yaml.bak-v0142-20260512T155012Z` plus matching profile backups.
- Hermes hook doctor is green across default, `personal-oss`, `earlypay`, and `infra-admin` profiles after `--accept-hooks` smoke.
- Fresh G4 report directory retained: `/Users/reddit/.agent-memory/reports/g4-v0138-20260512-132253/`.
- Fresh linkage diagnosis retained from G4 diagnostics: `g4-linkage-gap-diagnose-v0138-fresh.json` passed with decision `fresh_trace_linkage_gap_not_detected`.
- Fresh epoch readiness retained: `fresh-epoch-v0138.json` passed with decision `fresh_epoch_ready_to_compare_against_historical`.
- Fresh review queue preview retained: `g4-review-queue-preview-v0138-fresh.json` passed with decision `review_queue_ready_for_manual_review`, `read_only=true`, and `mutated=false`.
- G5a/G5b/G5c/G5d source checkpoint: `dogfood trace-cluster-preview`, `dogfood trace-candidate-persist/list/update/apply`, read-only trace-cluster scoring, and `dogfood reinforcement-refinement-preview` are merged/released through v0.1.141.
- G5d is merged/released via PR #302 and v0.1.141: repeated activation -> reinforcement refinement preview emits read-only/ref-safe candidates with `review_score` and `review_recommendation`; it writes JSON reports only, keeps `mutated=false`, and does not increment reinforcement counts, persist review state, promote memories, auto-approve ordinary conversation, or change retrieval defaults.
- G5a/G5b/G5c/G5d/G5e source checkpoint: `dogfood trace-cluster-preview`, `dogfood trace-candidate-persist/list/update/apply`, read-only trace-cluster scoring, `dogfood reinforcement-refinement-preview`, and `dogfood decay-collapse-preview` are merged/released through v0.1.142.
- G5e is merged/released via PR #306 and v0.1.142: stale weak evidence -> decay/collapse candidate preview emits read-only/ref-safe candidates and guardrails; it writes JSON reports only, keeps `mutated=false`, and does not persist review state, delete/deprecate/collapse memories, auto-approve ordinary conversation, or change retrieval defaults.
- Historical scheduled dry-run still blocks broad G4/background apply on `trace_quality_needs_more_dogfooding`, `decay_risk_above_threshold`, and `background_quality_warnings_present`.
- Broad G4/background apply remains blocked until the contract, historical reconciliation, narrow reviewed apply, decay/collapse, conflict/supersession, ranking eval, and rollback runway are verified.

Progress estimate:

- Overall north-star: 60-62%.
- Substrate/evidence plumbing: about 74-76%.
- Safe automatic mutation/promotion: about 42-45%.
- Remaining work: about 38-40% overall.
- Overall north-star: 62-64%.
- Substrate/evidence plumbing: about 75-77%.
- Safe automatic mutation/promotion: about 43-46%.
- Remaining work: about 36-38% overall.

Current interpretation:

- The fresh hook/runtime linkage blocker is resolved for v0.1.138-v0.1.141-era evidence.
- G5d completes another review-first brain-like signal loop, but it is not approval for automatic memory creation or reinforcement mutation.
- Broad G4/background apply remains blocked; fresh readiness, reviewed candidate apply support, G5c scoring, and G5d reinforcement-refinement preview do not authorize automatic memory creation.
- The next safe sequence is stale weak evidence -> decay/collapse candidate preview as review/preview-first G5e work; keep G4 broad apply and historical reconciliation as separate guarded corridors.
- The fresh hook/runtime linkage blocker is resolved for v0.1.138-v0.1.142-era evidence.
- G5e completes another review-first brain-like lifecycle signal loop, but it is not approval for automatic memory creation, decay/collapse mutation, or reinforcement mutation.
- Broad G4/background apply remains blocked; fresh readiness, reviewed candidate apply support, G5c scoring, G5d reinforcement-refinement preview, and G5e decay-collapse preview do not authorize automatic memory creation.
- The next safe sequence is conflict -> supersession/replacement candidate preview as review/preview-first work; keep G4 broad apply and historical reconciliation as separate guarded corridors.
- Existing broad-G4 baseline remains a docs/RED-test-only guardrail; do not advertise broad G4 consolidation apply mode as ready.

Current safe mutation boundaries:
Expand All @@ -52,8 +52,8 @@ Brain-like next design axis:
- `candidate -> reviewed fact/procedure/preference promotion` is available only through explicit G5b review/apply commands.
- `trace cluster -> review-priority scoring` is released G5c and remains human-review-only.
- `repeated activation -> reinforcement refinement preview` is released G5d and remains human-review-only; preview scores are not apply approval.
- Next: stale weak evidence -> decay/collapse candidate preview, read-only/ref-safe.
- Later: conflict -> supersession review.
- `stale weak evidence -> decay/collapse candidate preview` is released G5e and remains human-review-only; candidates are not delete/deprecate/collapse approval.
- Next: conflict -> supersession/replacement candidate preview, read-only/ref-safe.
- Retrieval ranking changes only behind opt-in eval before any default change.

---
Expand All @@ -77,30 +77,30 @@ For prompts like "다음으로 뭐해야 해?" or "다음 할 거 추천해줘",

- `.dev/status/next-agent-memory-action.md`

Current recommendation: start the next G5d repeated activation -> reinforcement refinement slice as review/preview-first work. Broad G4/background apply remains blocked by historical scheduled-dry-run debt and must not be enabled from a generic continuation prompt.
Current recommendation: start the next conflict -> supersession/replacement candidate preview slice as review/preview-first work. Broad G4/background apply remains blocked by historical scheduled-dry-run debt and must not be enabled from a generic continuation prompt.

## Ready-to-say answer

agent-memory is currently released/runtime-verified through `v0.1.140`. The installed Hermes hooks point at `/Users/reddit/.agent-memory/runtime/v0.1.140/.venv/bin/agent-memory`; package smoke reports `agent_memory.__version__ == 0.1.140`. G5a/G5b/G5c are merged/released for ref-safe trace-cluster preview, explicit reviewed trace-candidate persist/list/update/apply, and read-only `review_score`/`review_recommendation` signals for ref-safe clusters.
agent-memory is currently released/runtime-verified through `v0.1.142`. The installed Hermes hooks point at `/Users/reddit/.agent-memory/runtime/v0.1.142/.venv/bin/agent-memory`; package smoke reports `agent_memory.__version__ == 0.1.142`. G5a-G5e are merged/released for ref-safe trace-cluster preview, explicit reviewed trace-candidate persist/list/update/apply, read-only `review_score`/`review_recommendation` signals, repeated activation -> reinforcement refinement preview, and stale weak evidence -> decay/collapse candidate preview.

Broad G4/background consolidation apply mode remains blocked. Fresh linkage diagnostics passed, but historical scheduled-dry-run still blocks on `trace_quality_needs_more_dogfooding`, `decay_risk_above_threshold`, and `background_quality_warnings_present`. The first mutating review-queue/candidate corridors remain deliberately narrow and require explicit policy, approval phrase, actor, reason hash, backup, audit row, and rollback hint.

Historical G4 contract checkpoint remains docs/RED-test-only: PR #200, PR #202, PR #204, v0.1.99 runtime `/Users/reddit/.agent-memory/runtime/v0.1.99/.venv/bin/agent-memory`, and report `/Users/reddit/.agent-memory/reports/v0.1.99-runtime-qa-20260507T074118` are retained as the broad-G4-blocked baseline. Later releases hardened only narrow cleanup/restore/audit safety corridors, blocker diagnostics, and future trace/observation quality; they did not enable broad background consolidation mutation.
Historical G4 contract checkpoint remains docs/RED-test-only: PR #200, PR #202, PR #204, v0.1.99 runtime `/Users/reddit/.agent-memory/runtime/v0.1.99/.venv/bin/agent-memory`, and report `/Users/reddit/.agent-memory/reports/v0.1.99-runtime-qa-20260507T074118` are retained as the broad-G4-blocked baseline. Later releases hardened only narrow cleanup/restore/audit safety corridors, blocker diagnostics, fresh linkage, reviewed candidates, reinforcement review signals, and decay/collapse review signals; they did not enable broad background consolidation mutation.

## Current next slice

Completed release baseline: G4 fresh linkage/mutation safety landed by v0.1.136; G5a/G5b/G5c reviewed-candidate/scoring runway is released through v0.1.140; current active slice is the next G5d repeated activation -> reinforcement refinement, review/preview-first.
Completed release baseline: G4 fresh linkage/mutation safety landed by v0.1.136; G5a-G5e reviewed-candidate/scoring/reinforcement/decay preview runway is released through v0.1.142; current active slice is the next conflict -> supersession/replacement candidate preview, review/preview-first.

Current slice status: v0.1.140 is installed and live-smoked. G5c is complete and still read-only: it improves review prioritization for trace clusters without enabling broad background consolidation apply, telemetry reset apply, ordinary conversation auto-approval, raw transcript/query storage, or default retrieval ranking changes.
Current slice status: v0.1.142 is installed and live-smoked. G5e is complete and still read-only: it improves review prioritization for stale weak evidence and decay/collapse candidates without enabling broad background consolidation apply, telemetry reset apply, decay/delete/collapse mutation, ordinary conversation auto-approval, raw transcript/query storage, or default retrieval ranking changes.

Target shape for this slice:
Target shape for the next slice:

- `agent-memory dogfood fresh-epoch <db> --epoch-start <ISO>` emits `dogfood_fresh_epoch_readiness` with `read_only=true`, `mutated=false`, and `automation_policy.apply_supported=false`.
- The report excludes historical rows by timestamp and reports only aggregate counts for observation/trace/activation coverage, empty retrieval outcomes, trace distributions, candidate signals, and historical rows excluded.
- `agent-memory dogfood supersession-preview <db>` or equivalent emits a read-only/ref-safe conflict/supersession candidate report with `read_only=true`, `mutated=false`, `default_retrieval_unchanged=true`, and `automation_policy.apply_supported=false`.
- The report identifies same-claim-slot conflicts, replacement/supersedes chains, lifecycle status context, and copy-paste review commands using refs and aggregate counts only.
- No raw prompt/query/transcript/trace summary/sample values are printed.
- The first live source smoke against `/Users/reddit/.agent-memory/memory.db` with epoch `2026-05-09T21:57:33Z` wrote `/tmp/agent-memory-fresh-epoch-v0128-source.json`: 21 observations, 21 traces, 21 activations, coverage ratio 0.2381, 10 empty retrievals, blocked by `low_epoch_observation_trace_coverage` and `epoch_empty_retrieval_outcome_unknown`; no mutation.
- The first live G5e smoke against `/Users/reddit/.agent-memory/memory.db` wrote `/Users/reddit/.agent-memory/runtime/v0.1.142/g5e-live-smoke.json`: `read_only=true`, `mutated=false`, default retrieval unchanged, candidate count `0`, blocked only by `no_decay_collapse_candidates_ready`; no mutation.

Next safe slice: continue repeated activation -> reinforcement as preview/review-first G5d work. Do not live-apply queue/candidate mutations without an explicit operator decision and the exact guarded command shape. Broad G4 apply remains a separate, still-blocked slice.
Next safe slice: continue conflict -> supersession/replacement as preview/review-first work. Do not live-apply queue/candidate mutations without an explicit operator decision and the exact guarded command shape. Broad G4 apply remains a separate, still-blocked slice.

Recommended local backup commands before any future live mutation:

Expand Down
Loading