feat: Hermes scope and governed skills

.planning/MEMORYMASTER-DREAMING-V1.md

@@ -1,10 +1,11 @@
+<!-- doc-head: native Dreaming contract; Gemini extraction and GLM consolidation -->
 # MemoryMaster Native Dreaming V1
-> Covers: quiet transcript capture, asynchronous LLM consolidation, governed candidate writes, rollout, measurement, and rollback.
-> Key terms: Codex, Claude, OpenCode OAuth, GPT-5.6 Terra, exact evidence spans, capture ledger, candidate-first.
-> Read this before enabling Dreaming hooks, scheduling the worker, changing provider models, or activating candidate writes.
-> Default safety posture: disabled until explicitly installed; shadow processing before activation; never auto-confirms claims.
-> Authority: the claims store remains authoritative; the auxiliary capture ledger is replay state, not a second memory database.
-> Status: CURRENT implementation and stabilization contract; replacement 24-hour observation passed.
+# Covers: quiet capture, asynchronous consolidation, governed writes, rollout, measurement, and rollback.
+# Key terms: Gemini, GLM, exact evidence spans, capture ledger, candidate-first, task-bound providers.
+# Read when: enabling Dreaming, scheduling the worker, changing provider models, or activating writes.
+# Authority: claims remain authoritative; the auxiliary capture ledger is replay state, not a memory database.
+# Status: CURRENT; task-bound providers prevent ambient configuration drift and false-success results.
+<!-- /doc-head -->
 
 ## Intent
 
@@ -82,20 +83,19 @@ including schema-rejection paths, so hourly runs do not accumulate a second
 transcript archive. OpenCode credentials remain owned by OpenCode and are never
 read, copied, logged, or persisted by MemoryMaster.
 
-The local vNext stabilization uses ChatGPT OAuth for both stages:
-`openai/gpt-5.6-terra` at medium effort extracts typed evidence-linked
-candidates, while `openai/gpt-5.6-luna` at low effort performs the harder
-lifecycle comparison. GPT-5.4 Mini was removed from the local extraction
-configuration after live runs reproduced exit failures, malformed JSON, and
-low exact-evidence yield. Terra remains separate from Luna so model-specific
-stage budgets cannot consume one another. These are local activation choices;
-the portable extractor default remains Gemini.
+The active local configuration uses `gemini-3.5-flash-lite` for typed,
+evidence-linked extraction and `zai-coding-plan/glm-5.2` for lifecycle
+comparison. The scheduled action embeds both selections, clears stale variants,
+and runs Python in isolated mode, so ambient user variables cannot silently
+replace the chosen pair. Capture errors make the task fail nonzero even if the
+separate Dreaming phase has no errors. The portable extractor default remains
+Gemini and the portable consolidator default remains GLM.
 
 Verify account readiness without exposing credentials:
 
 ```powershell
 opencode auth list
-opencode models openai | Select-String 'openai/gpt-5.6-luna'
+opencode models zai-coding-plan | Select-String 'zai-coding-plan/glm-5.2'
 ```
 
 The scheduled task must run as the same Windows user that authenticated OpenCode. Missing CLI/account/model availability produces an actionable, retryable failure; it never silently switches providers.

.planning/PAPER-RADAR-REVIEW-2026-08-08.md

@@ -0,0 +1,216 @@
+# MemoryMaster paper-radar review - 2026-08-08
+# Covers: primary-paper findings, current-system gaps, and bounded implementation decisions for MemoryMaster.
+# Key terms: paper radar, filesystem memory, BudgetMem, A2RAG, temporal memory, governed skills, Mem2ActBench.
+# Read when: selecting research-derived work or checking why a memory technique was adopted, benchmarked, or rejected.
+# Sources: VoltAgent radar commit c8502b6 plus 18 version-pinned arXiv PDFs; upstream summaries are discovery only.
+# Verdict: adopt bounded capabilities, benchmark seven hypotheses, retain existing boundaries, reject wholesale rewrites.
+# Updated: 2026-08-08 after metadata triage, primary-PDF extraction, result/limitation review, and figure inspection.
+
+## Executive verdict
+
+The papers do not justify replacing MemoryMaster. They reinforce its strongest
+choices: authoritative governed claims, preserved evidence, explicit scope,
+lifecycle state, human promotion, and graph results rehydrated through claims.
+They also expose five useful gaps:
+
+1. LongMemEval retrieval and QA do not prove that an agent can apply memory to
+   a tool action without hallucinating missing parameters.
+2. Graph recall rehydrates claims, but it lacks a bounded fallback that maps a
+   graph/claim signal back to exact evidence excerpts when extraction omitted a
+   qualifier.
+3. `event_time`, validity intervals, and supersession exist, but temporal recall
+   still ranks mostly by freshness instead of query time, occurrence time, and
+   durative state.
+4. `personal-skill-v1` has governance, versions, validation, and citations, but
+   not explicit execution outcomes or failure-derived warnings.
+5. Token budgets exist, but cost is not attributed consistently to retrieval,
+   extraction, graph expansion, evidence admission, skill review, and answer
+   generation stages.
+
+The first implementation wave must add evaluation and telemetry before changing
+retrieval. No paper-derived behavior becomes a default without a reproducible
+quality or cost win on MemoryMaster's authoritative path.
+
+## Review scope and evidence standard
+
+- Discovery feed: `VoltAgent/awesome-ai-agent-papers` at commit
+  `c8502b6acd3978a84b8b25453eda24be83088d00`.
+- The pinned Memory & RAG section contains 57 parseable paper records although
+  its heading says 56 and its table of contents says 57. Parser output, not the
+  displayed count, is authoritative for the radar snapshot.
+- All 57 Memory & RAG records received metadata and abstract-level triage using
+  primary arXiv metadata.
+- Fifteen high-relevance Memory & RAG PDFs, the earlier filesystem-memory PDF,
+  and two cross-section evaluation/cost PDFs received a method, result,
+  ablation, limitation, and applicability review.
+- Key pages from the filesystem, BudgetMem, Skill-Pro, A2RAG, temporal-memory,
+  and Mem2ActBench papers were rendered and visually inspected to verify tables
+  and diagrams rather than relying only on extracted text.
+- Upstream descriptions and paper claims are not accepted as facts about
+  MemoryMaster. Every proposal below is compared with current code and must be
+  tested locally.
+
+## Current MemoryMaster baseline
+
+| Surface | Already present | Research-exposed gap |
+|---|---|---|
+| Retrieval | Explicit profiles, query classification, score explanations, token-bounded packing | Profiles mostly change weights; no measured multi-stage evidence sufficiency or admission policy |
+| Scope | SQLite and graph queries filter authorized scopes before trusted results are returned | Add invariant tests proving every future routing stage masks unauthorized data before scoring or provider calls |
+| Evidence | Source -> evidence -> claim lineage with content hashes and citations | No query-time exact-evidence map-back when a claim/edge lacks a required qualifier |
+| Graph | Supported edges, confirmed active claim authority, scope filtering, replay-safe support, claim rehydration | No progressive local -> path -> evidence fallback with explicit sufficiency diagnostics |
+| Temporal | `event_time`, `valid_from`, `valid_until`, supersession, freshness profile | No occurrence-time intent matching, interval overlap, or durative-state projection |
+| Skills | Strict `personal-skill-v1`, activation cues, workflow, validation, immutable versions, citations, human promotion | No success/failure/ambiguous execution evidence, termination condition, or negative warning path |
+| Evaluation | LongMemEval retrieval/QA, capture/graph quality, latency, replay and scope gates | No active tool-use benchmark, preservation score, or per-stage cost attribution |
+
+## Paper decisions
+
+| Paper | Verdict | What MemoryMaster should take | What it should not take |
+|---|---|---|---|
+| [Filesystem-Based Memory for LLM Agents](https://arxiv.org/abs/2607.26637v1) | **Adopt evaluation** | Measure answer quality, preservation, store health, consumer strength, and total retrieval cost separately | Filesystem authority, autonomous reorganization, or organization as a quality proxy |
+| [BudgetMem](https://arxiv.org/abs/2602.06025v1) | **Benchmark** | Explicit low/mid/high query budgets and stage-level cost/quality frontiers | An RL router or multiple model tiers before deterministic policies beat the baseline |
+| [Skill-Pro / ProcMEM](https://arxiv.org/abs/2602.01869v1) | **Adopt bounded schema ideas** | Activation, execution, termination, outcome evidence, and validation before reuse | Autonomous PPO evolution, score-based deletion, or automatic promotion |
+| [E-mem](https://arxiv.org/abs/2601.21714v1) | **Benchmark** | Bounded reconstruction from preserved contiguous evidence for multi-hop/narrative queries | Multiple resident memory agents or unbounded uncompressed contexts |
+| [ShardMemo](https://arxiv.org/abs/2601.21545v1) | **Retain and harden** | Scope-before-routing, cheap-first tiers, versioned skills, safe fallback to evidence | Learned MoE sharding for a personal SQLite corpus without measured scale pressure |
+| [A2RAG](https://arxiv.org/abs/2601.21162v1) | **Adopt** | Progressive local/path expansion, evidence-sufficiency diagnostics, and exact provenance map-back | Graph-only answers, unrestricted retry loops, or PPR before simpler traversal is measured |
+| [Less is More for RAG](https://arxiv.org/abs/2601.17532v1) | **Benchmark** | Admission control, redundancy/conflict pruning, pass-rate and drift telemetry | Treating uncertainty reduction as truth or adding an LLM probe to every recall by default |
+| [Grounding Agent Memory in Contextual Intent](https://arxiv.org/abs/2601.10702v1) | **Benchmark** | Explicit goal/action/entity cues and coarse episode boundaries for opt-in task retrieval | Multiple ingestion LLM calls per turn or uncontrolled label evolution |
+| [Beyond Dialogue Time](https://arxiv.org/abs/2601.07468v1) | **Adopt** | Occurrence-time retrieval, interval overlap, and a derived durative-state projection | Rewriting atomic evidence into a new authority or fixed monthly granularity |
+| [Reliable Graph-RAG for Codebases](https://arxiv.org/abs/2601.08773v1) | **Retain boundary** | Prefer deterministic structural providers and measure indexing coverage | A second MemoryMaster code graph; GitNexus remains the code-topology specialist |
+| [Seeing through the Conflict](https://arxiv.org/abs/2601.06842v1) | **Adopt observability only** | Separate semantic relevance, evidence consistency, and answer sufficiency in diagnostics | Trusting model-parametric memory over confirmed evidence or learned soft prompts |
+| [Amory](https://arxiv.org/abs/2601.06282v1) | **Benchmark** | Narrative-arc co-retrieval from preserved evidence | Autonomous narrative rewriting or synthetic-conversation conclusions as production proof |
+| [Controllable Memory Usage](https://arxiv.org/abs/2601.05107v1) | **Benchmark later** | An explicit caller-selected memory-reliance profile for fresh-start versus continuity tasks | Silent inference of how much history to obey or model fine-tuning for the first version |
+| [Proactive Memory Extraction](https://arxiv.org/abs/2601.04463v1) | **Benchmark** | Targeted re-extraction when a query exposes missing evidence; track integrity separately from accuracy | Repeated self-questioning on every capture or automatic replacement of preserved evidence |
+| [Membox](https://arxiv.org/abs/2601.03785v2) | **Adopt deterministic subset** | Retrieve adjacent evidence spans and link recurring source episodes without rewriting them | LLM-created topic boxes as authoritative memory |
+| [MAGMA](https://arxiv.org/abs/2601.03236v1) | **Defer architecture** | Compare temporal, causal, and entity path signals inside the supported graph | Parallel semantic/temporal/causal graph authorities or policy-learned traversal |
+| [Mem2ActBench](https://arxiv.org/abs/2601.19935v1) | **Adopt benchmark shape** | Score retrieval miss, retrieved-but-unused, hallucinated default, lossless-retention failure, tool error, and exact argument grounding | Treat its synthetic offline tool calls as sufficient production proof |
+| [Tokenomics](https://arxiv.org/abs/2601.14470v1) | **Adopt telemetry** | Attribute input/output/reasoning tokens and latency by MemoryMaster stage | Generalize its 30-task, one-framework results as MemoryMaster's expected distribution |
+
+## Ordered implementation packages
+
+### PPR-1 - Representation and active-use evaluation
+
+Add a versioned evaluation set and harness before product behavior changes:
+
+- latest versus superseded state;
+- occurrence time versus dialogue time;
+- valid interval and durative-state questions;
+- affect/emphasis preservation;
+- narrative-arc co-retrieval;
+- active tool invocation with exact parameters;
+- missing/default/inferred parameter distinctions;
+- retrieval miss, retrieved-but-unused, hallucinated default, lossless-retention
+  failure, and wrong-tool attribution;
+- answer correctness and citation correctness scored independently.
+
+Run the matrix against claims-only, evidence-only, claims+evidence,
+claims+approved-skills, and claims+ephemeral-guidance. Retain the existing
+LongMemEval R@5/MRR and full-QA regression gates.
+
+### PPR-2 - Stage-level sustainability telemetry
+
+Emit bounded per-request stage observations for retrieval, graph expansion,
+evidence map-back, admission, packing, skill recall/review, and answer/judge
+generation. Record elapsed time, provider calls, content read, input/output
+tokens when available, cache state, selected tier, fallback reason, and final
+correctness in evaluation artifacts. Do not persist private query or evidence
+text in aggregate telemetry.
+
+### PPR-3 - Deterministic budget and admission policy
+
+Introduce an explicit versioned policy selected by the caller:
+
+- `low`: lexical/confirmed claims, small evidence budget, no provider call;
+- `balanced`: current governed recall plus bounded graph/evidence fallback;
+- `high`: larger candidate window and explicit evidence-sufficiency check;
+- `temporal`: lifecycle timeline, occurrence-time/interval matching, evidence;
+- `procedural`: confirmed skills, warnings, and supporting claims.
+
+Start with deterministic rules and shadow evaluation. Add redundancy,
+near-duplicate, lifecycle-conflict, and weak-support admission diagnostics before
+testing any generator-aligned LLM pruning. Scope and sensitivity filtering must
+occur before tier selection, scoring, provider access, or cache lookup.
+
+### PPR-4 - Progressive claim-to-evidence rehydration
+
+Treat graph/entity matches only as navigation signals:
+
+1. retrieve authorized confirmed claims;
+2. expand one bounded supported path when the query is relational;
+3. test whether required entities, relations, temporal qualifiers, and citations
+   are present;
+4. map selected claims through `claim_evidence_links` to exact evidence excerpts;
+5. return a diagnostic fallback reason when evidence remains insufficient.
+
+No graph-generated fact may bypass claim status, scope, sensitivity, retired
+source, or citation checks.
+
+### PPR-5 - Temporal and episode projections
+
+Add derived, rebuildable projections over authoritative claims/evidence:
+
+- query-time versus occurrence-time intent;
+- interval overlap using `valid_from`/`valid_until`;
+- explicit latest/current versus historical selection;
+- bounded adjacent evidence windows from source order;
+- recurring episode links derived from stable source/session metadata;
+- durative state summaries that cite every contributing claim and never replace
+  atomic history.
+
+This package needs temporal precision and supersession adversarial tests before
+any schema or ranking change.
+
+### PPR-6 - Outcome-aware governed skills
+
+Extend skill evidence additively with execution observations:
+
+- outcome: `success`, `failure`, or `ambiguous`;
+- consumer/model profile and tool/schema snapshot;
+- activation match, termination result, validation result, and bounded metrics;
+- failure-derived warnings kept separate from positive procedures.
+
+Success may strengthen a review signal; failure must not strengthen a positive
+skill. No outcome automatically confirms, rewrites, prunes, or archives a skill.
+The steward and operator remain the only promotion authority.
+
+## Remaining radar triage
+
+The other 42 Memory & RAG papers remain in the radar, not discarded:
+
+- training-heavy routers, reinforcement-learning retrieval, and learned memory
+  controllers are deferred until deterministic policies have a measured ceiling;
+- domain-specific financial, supply-chain, scientific, SOP, embodied, multimodal,
+  and text-table systems are reference material unless MemoryMaster acquires the
+  corresponding use case;
+- surveys inform terminology but cannot establish an implementation gain;
+- multi-agent memory managers, autonomous compaction, learned forgetting, and
+  self-rewriting memory require adversarial preservation and governance evidence;
+- alternative GraphRAG systems remain benchmarks; they do not create another
+  authoritative graph or answer path.
+
+The next radar refresh should diff the pinned revision, classify only new or
+changed entries, and choose the next full-PDF batch from untested gaps rather
+than repeatedly rereading similar architectures.
+
+## Non-adoptions fixed by this review
+
+- No filesystem or Obsidian authority.
+- No automatic deletion, compression, consolidation, or trusted promotion.
+- No autonomous PPO/RL router in the personal SQLite profile.
+- No graph-only or vector-only answer authority.
+- No bulk paper-PDF ingestion into governed memory.
+- No full-context or multi-agent memory process as a default retrieval path.
+- No claimed improvement from paper-reported metrics without a MemoryMaster
+  baseline, mutation-relevant comparison, and reproducible local result.
+
+## Exit evidence required
+
+A research-derived package is complete only when its artifact records:
+
+- the exact dataset and provider/model identity;
+- the unchanged baseline and the candidate run;
+- quality, preservation, citation, cost, latency, scope, and replay results;
+- a wiring-relevant negative control or mutation where applicable;
+- explicit `adopt`, `retain`, `defer`, or `reject` disposition;
+- code and rollback scope;
+- zero secret and cross-scope leakage;
+- no regression beyond the existing LongMemEval and full-QA thresholds.