…ia infer=false

Replace passive auto-capture with skills-based memory extraction.
The agent evaluates each conversation turn and stores facts directly
to mem0 with infer=false, bypassing the extraction LLM entirely.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

make install — checks deps, installs OpenClaw if missing, builds plugin,
removes existing version, links local build, patches openclaw.json with
all required settings (tools.profile, session-memory, skills config),
backs up original config, and restarts gateway.

Also: make status, make logs, make restart, make uninstall, make clean.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Configure must run AFTER plugin link (link overwrites openclaw.json).
Moved Python config logic to scripts/configure.py to avoid Make shell
escaping issues with inline Python.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

The agent now rewrites search queries for retrieval instead of passing
raw user messages. The skill protocol teaches the agent to extract key
concepts, remove conversational framing, and match the language of
stored memories.

Plugin-side sanitization (metadata stripping) remains for the auto-recall
path, but all semantic query rewriting is the agent's responsibility.

Updated: skill-loader.ts (inline protocol), recall-protocol.md (examples),
recall.ts (comments clarifying responsibility boundary).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Rewrote recall-protocol.md with structured 4-step rewriting process,
7 worked examples with explicit reasoning chains, anti-pattern catalog,
and query length guidance. No ambiguity, no room for improvisation.

Inline protocol (prependSystemContext) updated with WRONG/RIGHT pairs
and compressed 4-step process for per-turn injection.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

lastCleanUserMessage was process-global mutable state written by
message_received and consumed by before_prompt_build. Under overlapping
inbound turns, one session's clean message could shape another session's
recall query, injecting wrong memories even with correct user_id scoping.

Fix: removed message_received hook and the shared variable entirely.
Query sanitization now runs inline within before_prompt_build where
ctx.sessionKey is available and execution is session-scoped. Uses
sanitizeQuery() (exported from recall.ts) to strip OpenClaw metadata.

The inherited currentSessionId shared state (line 95, pre-existing in
main branch v0.4.0) remains as a documented known limitation. That
requires OpenClaw framework changes (session context in tool handlers)
to fix properly.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…olicy

The facts array applied one category, importance, expiration_date, and
immutable flag to every entry in deduced_memories. Mixed-category batches
produced wrong retention policy: identity facts could inherit project TTLs,
or project memories could become permanently immutable.

Fix: batch by policy, not by turn. All facts in one memory_store call
must share the same category. If a turn produces facts in different
categories, the agent makes one call per category (e.g., identity facts
in one call, decision facts in another). Two categories = two API calls.
This is correct and expected.

Changes:
- Tool schema: facts[] description states same-category requirement
- Tool schema: category description explains it determines retention policy
- Tool schema: importance marked as override (omit to use category default)
- Inline protocol: teaches "batch by category" with mixed-category example
  showing two separate calls

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

loadTriagePrompt() was a hardcoded inline protocol that ignored domain
overlays, category overrides, custom rules, and triage knobs from the
config surface. The plugin advertised a configurable skill system via
skills.domain, customRules, and categories, but the live
before_prompt_build path ran a separate prompt plane.

Fix: loadTriagePrompt() now calls loadSkill("memory-triage") as the
primary path, which merges all config-driven overlays (domain, categories,
custom rules, triage knobs). The inline protocol is now a minimal
fallback for when the SKILL.md file cannot be loaded.

The full SKILL.md content (with all overlays merged) goes into
prependSystemContext, which is provider-cacheable. Tool usage format,
batching rules, and search protocol are appended as operational sections.

One prompt plane. Config surface = what the model sees.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…ructions

Three fixes:

1. SKILL.md taught old memory_store(text, {metadata}) contract while
   runtime expects facts[] array batched by category. Updated all 12
   examples to use facts: [...], category: "..." format. Added Example 11
   showing mixed-category turn with separate calls.

2. recall.enabled=false disabled the search path in index.ts but the
   system prompt still included search/filter instructions. Now gated:
   search instructions only injected when recall.enabled !== false.

3. Removed "MAXIMUM 3 operations per turn" from SKILL.md (replaced
   with "BATCH BY CATEGORY" which is the actual contract).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…ime guard

1. Replaced every remaining memory_store("text") example in SKILL.md
   entity-grouping section (lines 113-136) and consolidation example
   (line 332) with facts: [...], category: "..." format. Zero legacy
   examples remain (verified: grep returns empty).

2. Added runtime warning when multi-fact batch arrives without a
   category. The prompt teaches batch-by-category but the runtime now
   logs when the model disobeys, making violations observable without
   silently applying wrong retention policy.

Note on memory_search references in SKILL.md: these appear in
update/consolidation examples where the agent needs to find an existing
memory before replacing it. This is correct even when recall.enabled=false
because that flag disables pre-turn auto-search, not the memory_search
tool itself. The agent can always search explicitly for update operations.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Subagents were told not to store memories via system prompt, but
memory_store and memory_forget tool handlers did not enforce it.
A disobedient subagent tool call could write to a transient namespace
that is never read again, or delete memories from the parent namespace.

Fix: both memory_store and memory_forget now check
isSubagentSession(currentSessionId) and return an error response
before any provider call. The block is logged for observability.

This closes the gap between prompt-level instruction ("do not store")
and runtime enforcement. memory_search, memory_get, and memory_list
remain available to subagents (read-only access is safe and useful
for subagent context).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…tools

Three new tools to close the gap with mem0 MCP:

- memory_update(memoryId, text): Atomic in-place update that preserves
  history. Preferred over delete-then-store for corrections.

- memory_delete_all(confirm, userId): Bulk deletion with safety gate
  (confirm: true required). For user-requested memory resets.

- memory_history(memoryId): View edit history of a memory showing all
  changes over time with old/new values and timestamps.

All three tools enforce subagent write blocks. Both Platform and OSS
providers implement the new methods. OSS history() degrades gracefully
if the backend does not support it.

Plugin now exposes 8 tools (was 5):
  memory_search, memory_store, memory_get, memory_list,
  memory_forget, memory_update, memory_delete_all, memory_history

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

The /new and /reset commands produce system prompts ("A new session was
started via /new or /reset. Run your Session Startup sequence...") that
were being sent to mem0 search as the recall query. This wasted API
calls and returned noise.

Fix: detect system/bootstrap prompts by content pattern and skip the
recall search. The memory protocol is still injected (prependSystemContext)
so the agent has the skill instructions, but no search fires for system
commands.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

skills.recall.strategy now drives real behavior:

- "smart" (default): long-term search only, no session search.
  1 plugin search per turn instead of 2. Agent handles additional
  search via memory_search tool.

- "manual": zero plugin searches. Agent controls all search.
  Skill protocol instructs the agent to search proactively at
  session start and when context is needed.

- "always": long-term + session search every turn (old behavior).
  2 plugin searches per turn.

Also: skip recall entirely for system/bootstrap prompts (/new, /reset).

This addresses the feedback that the system was more search-heavy than
the first-principles ideal of agent-controlled search. The default
"smart" mode halves the always-on search cost while keeping baseline
recall. "manual" mode gives full agent control.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…precedence

1. Skill protocol now teaches explicit scope selection:
   - "long-term" for user context (default, most common)
   - "session" for current conversation only
   - "all" only when both scopes truly needed
   Avoids unnecessary backend fan-out from scope="all".

2. Documented enabled/strategy precedence in types.ts:
   enabled=false overrides strategy. Strategy only consulted
   when enabled is not false.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

The skill protocol now explains user_id, agent_id, and run_id scoping:
- Default: plugin handles scoping automatically, agent does not need
  to pass userId or agentId in most cases
- agentId: only for cross-agent queries (accessing a different agent's
  memory namespace)
- userId: only when explicitly operating on a different user
- run_id: managed by the plugin via the scope parameter, not passed
  directly
- Multi-agent isolation: each agent has separate memory, main agent
  memories are separate from subagent memories

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Automatic memory consolidation that triggers based on activity:
- Three gates: time (24h), sessions (5), memory count (20)
- File-based lock prevents concurrent consolidation (1h stale)
- Piggybacks onto next user turn when gates pass
- 4-phase SKILL.md: Orient, Gather, Consolidate, Report
- Uses memory_update (atomic), memory_history, memory_search with filters
- State persists in plugin stateDir across gateway restarts

New: dream-gate.ts (gate logic, lock, state persistence)
Rewrite: skills/memory-dream/SKILL.md (4-phase structured prompt)
Modified: index.ts (wire gates into agent_end + before_prompt_build)
Modified: types.ts (dream.auto, minHours, minSessions, minMemories)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>