AgentMemory

### Changed

- Added the marketplace card media kit (hero and 3 illustrated panels from the launch film) and refreshed the listing description and features.

### Changed

- The marketplace listing shows the launch film. No skill change.

### Added

- Every AgentMemory route an agent token can reach now has a method (SP-40):
  `listMemories` and `getMemory`; the extraction sources (`listSources`,
  `getSource`, `createSource`, `updateSource`, `deleteSource`, `syncSource`);
  the session catalog (`listSessions`, `getSession`); the run histories
  (`listExtractionRuns`, `listMaintenanceRuns`); `getEntryStats` and
  `listEntryVerdicts`; the curator's `listMemoryVerdicts`;
  `getDeliveredMemory`, `exportMemories` and `listDeskMemories`; uploads
  (`ingestSessionEvents`, `syncKnowledge`); pipeline runs
  (`extractAgentSessions`, `triggerExtraction`, `runMaintenance`); and the
  profession reads `getProfession` and `listProfessionWorkflows`. Each names
  its reply's top-level fields. No route changed.
- Every method now takes every field its route takes (SP-40 review bar):
  `listMemories` and `listDeskMemories` gain `forgotten` (the desk list also
  `min_confidence`), `getStats` gains `agent_id`, `createMemory` gains
  `description` and `source_session_ids`, and `syncKnowledge`'s `truncated`
  names its `reason`. A field the gateway sets itself (`origin_brand_id`,
  `reviewer`, `resolved_anchors`, a desk's `agent_id`, the craft queue's
  `unjudged`, `version_pin`) is named as runtime-only in the method's
  description.

### Added

- Brand Memory Curators can list recent agent memories for shadow review and
  submit evidence-first keep, forget, correction, or pin proposals.

### Changed

- `entriesNear` now documents current guard paths, both truncation signals,
  and the fallback for older servers; `anchorsLookup` documents current guard
  paths, real counts, and its configurable per-node lesson limit.

### Added

- Typed-entry reads describe lesson guards and origins: `searchEntries` and
  `entryContext` expose `unguarded` and `derived_from` filters and document
  guard-aware result metadata, while `entriesNear` documents each entry's
  guard and its final ranking tie-break.

### Fixed

- `entriesNear.node_ids` description: the gateway DOES resolve file paths in
  this route — callers should use `node_ids` when they already hold them (for
  example from the code-index skill) and otherwise pass `paths`.

### Added

- `entryContext` accepts `author` and `stale` lifecycle filters alongside its
  code-project context filter.

### Added

- `entryContext` accepts `author` and `stale`, matching the filters exposed by
  its shared `GET /entries/search` route.

### Added

- `anchorsLookup` (`POST /api/v1/memory/anchors:lookup`) finds lessons for a list of code-node ids you already hold, pinned first. Resolves file paths to node ids first with the `code-index` skill — never pass paths.

### Changed

- `searchEntries` (`GET /api/v1/memory/entries`) returns the relevant few. With
  `q`, results are ranked by keyword match plus meaning, rows below a
  similarity cut-off (default 0.5) are dropped unless they match a keyword, and
  the default `limit` is 20 (maximum 500).
- With an empty `q` the call is a browse: up to 500 matching entries, pinned
  first, then newest. Callers that never send `limit` get the same 500 as before.

### Added

- `limit` query parameter on `searchEntries` and `entryContext` (1-500).

### Added

- `contributeSkillLesson` and `pullSkillLessons` share reviewed learnings and
  patterns through an installed skill's cross-engagement craft store.
- `listCraftPendingCodes`, `curatorCraftQueue`, and `postCraftVerdict` expose
  the probationary Memory Curator's shadow-review workflow without applying
  publication or rejection decisions.

### Added

- `listPendingEntries` and `submitEntryVerdict` expose the Memory Curator's required shadow-review workflow, including the pending route's required `since` query parameter.

### Added

- **`updateMemory` takes `description`** (≤200 chars), the one-line summary
  that recall returns beside a memory's title. 1.2.0 withheld it on purpose:
  the server accepted the field and then discarded it, because
  `update_memory` applied a fixed list of fields that did not include it, so
  a PATCH carrying it returned 200 and changed nothing. The AgentMemory
  service now applies it (AS-410). Before that fix is deployed, the field is
  still accepted and still ignored. A value over 200 characters is refused
  with a 422.

### Fixed

- `references/deduplicate-before-write.md` still said the skill did not
  expose `updateMemory`, which stopped being true in 1.2.0. It now says to
  correct the memory in place, or to archive it when a new memory replaces
  it.

### Added

- **`updateMemory`** (`PATCH /memories/{memory_id}`) — correct a memory in
  place: title, body, kind, tags, confidence, importance, or
  `status: archived` to retire it from recall while keeping it auditable.
- **`deleteMemory`** (`DELETE /memories/{memory_id}`) — soft-delete a memory
  written by mistake.

Both routes already existed and were already tenant-scoped at both hops
(gateway proxy → AgentMemory → `memory_service.get_memory(db, tenant_id,
memory_id)`, which 404s outside the caller's tenant). The schema was
withholding them: `_drift_ignore` classified edit-and-delete-by-id as
"admin dashboard moderation surface", which left an agent with **no way to
correct or remove anything it had written**. `createMemory` does not upsert,
so a seeding script run twice duplicated every row with no cleanup path, and
a junk row at confidence 0.5 sits above the 0.4 workspace-sync floor — it
reaches the agent every hour for the ~60 weeks that 2%/week decay needs to
push it under the 0.15 archive threshold. Reported by `@di-atomic`
2026-09-21 while seeding memory for a client brand.

### Changed

- `SKILL.md` now states that `createMemory` does not upsert, that decay is
  not a cleanup mechanism, and that correcting beats writing a second
  contradicting memory.

> ⚠️ **Reconstructed 2026-09-21 from the published bundle, not written at release time.**
> 1.1.1 and 1.1.2 were both published to the marketplace and neither left an entry here, so
> this file read `1.2.0 → 1.1.0` while the registry served two versions in between. Source
> for each is named below; nothing here is inferred from a version number.
>
> ASP-245 (#1792) *did* write a 1.1.2 entry — into the published bundle, and into 7 of its 8
> packages' repo changelogs. This one was lost to a collision: #1786 edited this same file in
> the same window and merged first. Seven siblings landing cleanly is what makes it a dropped
> hunk rather than an omission.

### Changed

- `skills/agentmemory/schema.yaml`: `_drift_ignore.field_mismatches` rewritten to match the
  live AgentMemory OpenAPI (ASP-240) — it had exempted three fields that do not exist and
  missed three that do. Gate-only metadata; it never reaches a generated client.

No agent-visible change: `methods:` is byte-identical to 1.1.1.

*Source: the `CHANGELOG.md` inside `https://bundles.opvs.ai/opvs-ai/agentmemory/1.1.2.tgz`,
which carries this text verbatim. Published 2026-09-21T09:29:53Z (ASP-245).*

### Fixed

- **PC-338 — four params were being sent in the JSON body, where the server never looks.**
  `searchMemories` and `recallRelevant` post to `/agents/{id}/context` and
  `/agents/{id}/context/relevant`, whose `query` / `kinds` / `limit` / `format` are
  FastAPI `Query()` parameters. Without an explicit `in: query`, the generator infers
  placement from the verb and puts them in the body. The two routes then failed in
  **opposite** ways, which is why only half of it was reported: `/context/relevant`
  declares `query` required, so a body-placed value 422'd loudly on every call, while
  `/context` declares it optional, so `query` arrived as `None` and the route returned the
  agent's memories **unfiltered, with a 200**. A loud failure was the lucky half.

*Source: diff of `https://bundles.opvs.ai/opvs-ai/agentmemory/1.1.0.tgz` against `1.1.1.tgz`
— eight added `in: query` lines plus the PC-338 comment blocks, and nothing else but the
version string. Gate: `tests/api/skills/test_wire_placement_agentmemory.py`.*

### Added

- **Initiative A — envelope contract (sei).** `skills/agentmemory/schema.yaml` bumped to version 1.1.0 with skill-level `intent_aliases`, `pitfalls`, and per-method `guidance` blocks carrying `next`, `dont`, `warn`, and `telemetry_signal_default`. Matches the Wave 2 baseline established by `@opvs-ai/agentboard@1.1.0` and `@opvs-ai/agentdocs@1.1.0`. The four methods (`searchMemories`, `recallRelevant`, `createMemory`, `getStats`) now ship OPVS-native guidance inline.
- **Initiative B — 4-layer format (s4l) with research-grounded content.** Three new directories ship inside `skills/agentmemory/`:
  - `references/` — 5 procedure docs (`decide-what-to-persist`, `recall-with-relevance`, `deduplicate-before-write`, `kind-and-importance`, `memory-vs-docs-vs-tasks`)
  - `learnings/` — 4 research-grounded Knowledge Items (write-discipline from Mem0/Letta/Anthropic, staleness-is-known-unsolved per Mem0 2026, recall-relevant-over-search, pii-and-secrets per Mem0 warning)
  - `registry/` — 1 typed asset (`engineering-agent-starter-memories` template) + `registry.json` index
- **Industry research cited** in references/ + learnings/: Mem0 docs + Mem0 "State of AI Agent Memory 2026" blog (hybrid retrieval, scopes, staleness as unsolved problem); Letta/MemGPT (3-tier core/recall/archival memory architecture); Anthropic memory tool ("assume interruption" multi-session pattern); Cognee (knowledge graph as complement to vector recall); ChromaDB/Pinecone production patterns (TTL, threshold tuning, metadata-rich storage).

### Changed

- Schema header version: `1.0.0` → `1.1.0` (was stale — package was already on 1.0.3).
- Top-level description tightened: "Long-term agent memory — search, recall, and persist decisions, learnings, patterns, preferences, and facts across sessions. Hybrid FTS + LLM-reranked retrieval; explicit writes, per-brand isolation, confidence decay."
- `searchMemories.description` clarified — hybrid FTS + tsvector + GIN ranking.
- `recallRelevant.description` clarified — two-stage FTS pre-filter (100) → LLM reranker (top N). Now distinguishes its cost profile vs `searchMemories`.
- `createMemory.body` no longer has a max_length cap in the schema (compression handles the prose path).

### Added

- **Marketing metadata** — `marketing` block now carries a `long_description` (2–4 paragraphs of admin-facing copy), 5–7 `features` bullets with lucide-react icons, and an `author` record ({name: OPVS Team, email: ops@opvs.ai, website: https://opvs.ai}). Drives the Wave 2 Phase B customer detail page at `/dashboard/marketplace/packages/@opvs-ai/<name>`. `media` block intentionally empty for this release — visual assets (hero image, screenshots, demo video) come in a later wave.

### Changed

- `tier: internal` → `tier: beta`. The dashboard customer browse at `/dashboard/marketplace` strips tier=internal by design (per marketplace concept: "internal = no public catalog entry"). Wave 1 is customer-facing — brand admins install `@opvs-ai/platform-skills` to get the OPVS platform. `beta` is the right tier per concept: "published + opted-in per brand, pre-1.0."
- Peer_dependency ranges still `^1.0.0` — matches 1.0.0, 1.0.1, AND 1.0.2 via npm-style.

### Changed

- Republished with `meta_coherence` fix (server-side 979b3b41) — score now 100/100. The 1.0.0 publish scored 95 due to a false-positive `meta.mixed_semantics` warning that penalised every package declaring `@opvs-ai/opvs-foundation` as a peer_dep. The warning has been removed; the BLOCKING `meta.empty_useless` guardrail stays.

### Added

- Initial marketplace release. Packages the in-tree `agentmemory@1.0.0` skill
  verbatim — search, create, and stats methods over the agent memory service.
- Peer dependency on `@opvs-ai/opvs-foundation@^1.0.0`.