Redesigned tool set around Nowledge Mem's v0.6 architecture.

Tools:
- Add nowledge_mem_save: structured capture with 8 unit_types
- Add nowledge_mem_context: Working Memory daily briefing
- Add nowledge_mem_connections: graph exploration with EVOLVES chains,
  entity discovery, cross-topic synthesis, and SOURCED_FROM provenance
  for Library documents/URLs
- Add nowledge_mem_forget: delete by ID or query with confirmation
- Remove nowledge_mem_search (redundant with memory_search)
- Remove nowledge_mem_store (replaced by nowledge_mem_save)
- Remove nowledge_mem_working_memory (replaced by nowledge_mem_context)

Tool descriptions sharpened based on live testing:
- connections: explicit cross-topic synthesis and provenance use cases
- save: proactive signal — don't wait to be asked
- memory_search: temporal query guidance

Hooks:
- recall: updated tool guidance to map user intents to right tools
- capture: multi-layer quality gate (questions, injection, triggers)
- after_compaction hook added for thread continuity

Slash commands: add /forget

Co-authored-by: Cursor <cursoragent@cursor.com>

- Add nowledge_mem_timeline: wraps /agent/feed/events with day-grouped
  output for temporal questions (last_n_days, event_type, tier1_only)
- Update memory_search description to reflect multi-signal scoring:
  embedding + BM25 + labels + graph/community + recency decay
- Update recall hook: route temporal queries to timeline, topic queries
  to memory_search/connections
- Update nowledge_mem_context: clarify WM is read-only from plugin side
- Document WM full-overwrite API and backend gaps in CLAUDE.md

Co-authored-by: Cursor <cursoragent@cursor.com>

- Add event_date_from/to and recorded_date_from/to params to memory_search
  Two time dimensions: when fact HAPPENED vs when it was SAVED
  Uses client.searchTemporal() via API directly (works with any CLI version)
- Results include eventStart, eventEnd, temporalContext from backend
- CLAUDE.md: document all API signatures with params

Co-authored-by: Cursor <cursoragent@cursor.com>

…meline chaining

- connections: rewrite output to JOIN edges→nodes; show CRYSTALLIZED_FROM,
  EVOLVES (with sub-relation labels), SOURCED_FROM, MENTIONS each as typed
  sections with relationship strength and memoryIds for follow-up
- memory_search: switch to searchRich (API-direct) — surface relevanceReason
  ("Text Match 100% + Semantic 69% | decay[imp:high]") and importance per result
- timeline: embed memoryId hints in entries so agent can chain to connections
- timeline: document event_type values in description for model discoverability
- recall hook: use searchRich — show scoring breakdown in injected context
- client: add searchRich() convenience wrapper

Co-authored-by: Cursor <cursoragent@cursor.com>

Aligns with nmem-cli v0.4 which adds graph expand (g) and feed (f) commands.
Every feature now works in remote mode via NMEM_API_URL + NMEM_API_KEY
without any plugin changes.

- client: graphExpand() uses `nmem g expand`, feedEvents() uses `nmem f`
- client: search() now gets rich fields (relevance_reason, importance, labels,
  temporal) from the enriched CLI JSON — searchRich() is now an alias
- client: searchTemporal() uses CLI --event-from/--recorded-from args
- client: _normalizeMemory() canonical shape across all search paths
- connections: uses client.graphExpand() instead of client.apiJson()
- timeline: uses client.feedEvents() instead of client.apiJson()
- All API fallbacks retained for graceful degradation on older CLI versions

Co-authored-by: Cursor <cursoragent@cursor.com>

…giene

Exposes Nowledge Mem remote access in plugin configuration:
- apiUrl: server URL (empty = local, supports remote deployment)
- apiKey: API key, marked secret=true in uiHints for masked UI input

Credential handling rules:
- apiKey injected ONLY as NMEM_API_KEY env var into child processes
- Never passed as a CLI arg (avoids ps aux exposure)
- Never logged at any level
- Config values win over NMEM_API_URL / NMEM_API_KEY env vars
- Initialization log shows mode=remote → URL vs mode=local

NowledgeMemClient now accepts credentials at construction; _spawnEnv()
and _apiUrlArgs() build per-spawn env and args cleanly.

Co-authored-by: Cursor <cursoragent@cursor.com>

Destructure error from both useCachedPromise calls and display a
Toast.Style.Failure when searchMemories or listMemories fails, so
users see "Connection failed" instead of a silent empty list.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…_type fixed

nowledge_mem_save now exposes the full save contract:
- unit_type: works end-to-end (API model fix in backend)
- labels[]: array of topic/project tags, passed as repeated -l args
- event_start / event_end: bi-temporal "when it happened" fields
- temporal_context: past | present | future | timeless
- importance: clamped to [0, 1]

Confirmation message shows all non-default fields:
  "Saved: DB choice [decision] (id: ...) · labels: backend · event: 2024-03"

CLAUDE.md: full rewrite reflecting current architecture
  - CLI-first table, remote config rules, tool surface, known gaps
README.md: remote mode config section with secure apiKey example

Co-authored-by: Cursor <cursoragent@cursor.com>

Gap 1 — Feed date range:
- nowledge_mem_timeline: add date_from/date_to (YYYY-MM-DD) params for exact
  temporal queries ("what was I doing last Tuesday?")
- client.feedEvents(): pass --from/--to to `nmem f`; API fallback uses
  ?date_from=&date_to= query params
- Range label in timeline output: "Activity 2026-02-17 → 2026-02-17:"

Gap 2 — EVOLVES version chain via CLI:
- client.graphEvolves(): new method using `nmem g evolves <id>`; falls back
  to GET /agent/evolves?memory_id=<id> (newly supported by backend)
- connections.js: switch from raw apiJson() to client.graphEvolves(); edge
  direction now shown relative to the queried memory (→ newer / ← older)

Gap 3 — Working Memory section-level edit:
- client.patchWorkingMemory(heading, { content, append }): new method using
  `nmem wm patch`; falls back to JS patchWmSection() + full PUT
- nowledge_mem_context: now accepts patch_section + patch_content/patch_append
  params; agents can update one WM section without overwriting the rest
- patchWmSection() JS helper mirrors the Python _wm_patch_content() logic

Bump to v0.2.7

Co-authored-by: Cursor <cursoragent@cursor.com>

Bug: buildAgentEndCaptureHandler gated the thread append behind
shouldCaptureAsMemory(), which checks for memory-trigger patterns in the
last user message. Sessions ending with a question or a command were
silently skipped — the thread was never saved to Nowledge Mem.

The before_reset handler was already correct (unconditional).

Fix: move appendOrCreateThread() before the shouldCaptureAsMemory check.
Thread capture now always runs on successful agent_end when messages exist.
The memory note remains gated — it's an additional signal, not the
prerequisite for saving the session thread.

Verified end-to-end:
- thread create --id + append --idempotency-key both work in installed CLI
- second append with same idempotency-key returns messages_added=0
- sessions ending with questions now get thread-captured correctly

Co-authored-by: Cursor <cursoragent@cursor.com>

Expand Hooks section in README to explain:
- autoRecall and autoCapture are plugin lifecycle hooks, not LLM decisions
- autoCapture does two independent things: unconditional thread transcript
  append and conditional memory note extraction
- Thread deduplication by external_id prevents duplicates across hook fires
- Quality gates apply only to memory notes, not thread transcripts
- What "context compaction" is and why capture happens at that point

Co-authored-by: Cursor <cursoragent@cursor.com>