feat: Priority 3 Memory Enhancements — Shared Scope, Entity Graph, Confidence Scoring, Proactive Injection

[helal-muneer]

Summary

Four advanced memory features to make OpenClaw agents significantly smarter:

H. Cross-Agent Shared Memory (shared: scope)

• New shared built-in scope automatically readable by all agents
• Write access restricted to dreaming engine or explicit memory_shared tool
• Optional auto-promote: memories accessed by 3+ agents get promoted to shared during dream cycle
• Config: scopes.shared.enabled (default: true), scopes.shared.autoPromote (default: false)

I. Entity Relationship Layer (src/entity-graph.ts)

• Regex-based entity extraction (people, projects, tools, locations, preferences, orgs)
• Relationship extraction (subject-predicate-object triples)
• In-memory graph with getRelated(), getEntityProfile(), getAllEntities()
• Noop variant when disabled — zero overhead
• Config: entityGraph.enabled (default: false)

J. Memory Confidence Scoring (src/confidence-tracker.ts)

• Tracks recall count and useful-recall count per memory
• Formula: confidence = usefulCount / max(recallCount, 1) * decayBoost
• Decay boost decreases on unhelpful recalls, recovers on useful signals
• Integrates with auto-recall pipeline (automatic tracking)
• New tool: memory_boost — manually boost a memory's confidence

K. Proactive Memory Injection (src/proactive-injector.ts)

• Three trigger types:
  1. Entity-based: pre-fetch related memories when user mentions a known entity
  2. Pattern triggers: inject memories matching configurable regex patterns
  3. Stale memory: random 5% check for memories not recalled in N days
• Max 1 injection per turn, dedup against auto-recall results
• Config: proactive.enabled (default: false), staleMemoryDays, entityPrefetch, patternTriggers

New Tools

• memory_entities — query entity graph (profile/related)
• memory_boost — manually boost memory confidence
• memory_shared — explicitly write to shared scope

Files Changed

• Created: src/entity-graph.ts, src/confidence-tracker.ts, src/proactive-injector.ts
• Modified: src/scopes.ts, src/tools.ts, index.ts, openclaw.plugin.json
• Tests updated: scope-access-undefined.test.mjs, clawteam-scope.test.mjs

Design Principles

• All features off by default — no breaking changes
• No new external dependencies
• No modification to existing functionality — purely additive
• Conservative defaults with granular opt-in config

[rwmjhb]

The individual capabilities here are interesting, but several are shipped in a non-functional or unsafe state that would affect users who enable them. Two must-fix items before this is ready.

Must fix

• Proactive injection is wired but never called: The ProactiveMemoryInjector is instantiated but its inject() method is never invoked in the request pipeline. Users who enable proactiveInjection: true will see no behavior change. The injection call needs to be hooked into the retrieval or response path to deliver the advertised feature.
• Entity graph leaks across scopes: EntityRelationshipGraph queries all memories regardless of scope, then returns related entities to any requesting agent. An agent in scope A can receive entity facts extracted from scope B's private memories. Scope filtering must be applied at graph construction or query time before this is safe to enable in multi-agent deployments.

High priority

• scopes.shared.enabled: false has no effect — the shared scope is active unconditionally. The config guard is missing.
• Shared scope write restriction is not enforced: the description says only the dreaming engine or memory_shared tool may write to shared, but memory_store can write there directly. memory_shared also bypasses safeguards that memory_store already enforces.
• Confidence scoring is tracked in-memory only — state is lost on restart. The comment implies metadata persistence, but none exists. Users will see confidence reset to default after every restart.
• getRelated(entity, 1) traverses 2 hops — depth semantics are off by one.
• autoPromote config field is defined in schema and types but has zero implementation. The PR description describes auto-promotion during the dream cycle, but no promotion logic appears in the diff — where does it live?

Other

• timestamp=0 entries are always treated as stale — is this intentional or a sentinel value that should be handled separately?
• Preference entity extraction stores the verb rather than the preference value (e.g. "loves" instead of "coffee").
• Three new source modules (519 lines) have zero test coverage.

Questions for author

• Is this part of a maintainer-approved roadmap? The Priority 3 / H-K numbering implies earlier PRs landed first — were those merged with explicit sign-off to continue the series?
• CI fails with cli-smoke.mjs: undefined !== 1 at line 316 — pre-existing on base, or introduced here? Needs root-cause before merge.