Proactive Research — your agent notices, decides, and researches overnight

The proactive research loop turns the living-memory observation layer into initiative: the agent notices a recurring pattern about you (from what, when, and how you interact), concludes it deserves attention, researches it on the web overnight, and presents a briefing next morning — with every step auditable back to the memories that triggered it.

 living-memory signals            nightly (maintenance)                overnight (cron agent)        next session
┌──────────────────────┐   ┌───────────────┐  ┌────────────────┐   ┌──────────────────────────┐   ┌─────────────┐
│ valence-stamped facts│ → │ insight-      │→ │ research-      │ → │ research-overnight job:   │ → │ 🔎 briefing │
│ routines, patterns,  │   │ synthesis     │  │ trigger        │   │ pick → web research →     │   │ block (once │
│ frustration signals, │   │ (LLM, ≤2/run) │  │ (no LLM, ≤1/n) │   │ research store (briefing) │   │ per session)│
│ goals, identity refl.│   │               │  │                │   │                           │   │             │
└──────────────────────┘   └───────────────┘  └────────────────┘   └──────────────────────────┘   └─────────────┘

The four steps

1. Insight synthesis (insight-synthesis, nightly LLM step): reads the last 7 days of person-signals — negative-valence memories, mined routines, frustration signals, behavioral patterns — plus two current-state (not windowed) sources: the user’s active goals (goal registry) and the assistant’s own durable identity reflections on the relationship (“what patterns define good partnership with the user?”). It writes at most 2 evidence-linked insight facts (“User repeatedly works past midnight and expresses fatigue — sleep procrastination may be hurting focus”; “User’s goal to ship the v2 API has stalled for 18 days despite being high priority”). The model can only cite numbered evidence references (F#/S# for facts/signals, G#/I# for goals/identity reflections) that map back to real ids; an insight citing an unknown reference is dropped, so hallucinated evidence chains are structurally impossible. Insights are ordinary decayable memories tagged needs-review, with topic one of wellbeing/productivity/growth/tooling/relationship/other — growth marks insights grounded mainly in the user’s stated ambitions. Goal/identity evidence text is frozen into the insight’s provenance at synthesis time (goal state can change before the research loop re-reads it days later), while fact/signal evidence is re-read live downstream. Both sources are optional and skip cleanly when goal stewardship or identity reflection is off.
2. Trigger policy (research-trigger, nightly, deterministic — no LLM): at most one insight per night graduates to research. Gates: importance ≥ 0.74 (salience ≥ 0.6), evidence count, topic blocklist plus a sensitive-text regex (credentials never leave the machine), and a 14-day per-topic cooldown. The pick is recorded as a research:<slug> queue fact whose provenance points at the insight — which points at the raw evidence.
3. Research executor (research-overnight OpenClaw cron job, default 03:30, isolated heavy-model agent session): runs openclaw hybrid-mem research pick --json, researches the topic with the session’s web tools, and stores ≤400 words via openclaw hybrid-mem research store — the single writer for briefings, which validates topic state, length, and http(s) sources, and records the full provenance chain including every source URL. Most nights it replies SKIPPED: no research topic and exits. No web tools configured ⇒ TOOLING_BLOCKED and no briefing from prior knowledge — a briefing without checkable sources is worse than none.
4. Morning delivery: the next session start injects a one-time “🔎 Overnight research briefing” block (headline + memory_recall hint; index-only exposure, so surfacing it never counts as a recall). Optionally, setting research.delivery { mode: "announce", channel, to } also pushes the agent’s summary to that channel when the overnight run finishes.

Why you can trust it

• Explainable by construction. Christoffer’s “I can’t describe exactly why it decided to research this” is answered here by walking provenance: briefing → queue fact → insight → evidence facts/signals. openclaw hybrid-mem research status shows the queue, briefings, and evidence counts.
• Single-writer storage. The overnight agent cannot write memories directly; only the validating research store CLI can, and web content is handled as untrusted data (the job message forbids following instructions found in fetched pages; briefing headlines are injection-sanitized again at delivery).
• Bounded cost. One heavy-model agent turn per night at most, one topic per night, 14-day topic cooldown, 20h re-run guard, and the whole loop off with one flag.

Configuration (research.*, all default-on except announce delivery)

Key	Default	Effect
research.enabled	true	Master gate: synthesis, trigger, cron job, and delivery.
research.schedule	"30 3 * * *"	Cron expression for the overnight agent.
research.insights.maxPerRun / windowDays / minEvidence	2 / 7 / 2	Synthesis caps. insights.model overrides the maintenance-tier model.
research.trigger.minImportance	0.74	Insight importance floor (importance = 0.5 + 0.4 × salience).
research.trigger.cooldownDays / maxPerNight	14 / 1	Re-research cooldown per topic; nightly budget.
research.trigger.topicBlocklist	["credential","secret","security"]	Topics never researched (a sensitive-text regex also always applies).
research.executor.maxBriefingChars / maxSources	4000 / 8	Briefing storage caps.
research.delivery.injectDays / maxBriefings	3 / 2	Session-start injection window and per-session cap (0 disables injection).
research.delivery.mode + channel + to	"none"	Set all three for an announce push (e.g. Telegram) when the overnight run completes; partial config safely maps to none.

Operations

• openclaw hybrid-mem verify --fix (or install/upgrade) installs the research-overnight cron job alongside the maintenance jobs; research.enabled: false gate-disables it (and re-enables when flipped back).
• The nightly steps run inside the standard maintenance orchestrator (openclaw hybrid-mem maintenance nightly); summaries appear in the maintenance logs (stored=… candidates=… evidence=…, picked=… cooldown=… blocklist=…).
• Everything the loop produces is visible in the Memory Graph (insights, queue facts, briefings, and their provenance edges) and manageable like any other memory.

Related docs

• LIVING-MEMORY.md — the observation layer this loop builds on
• OPERATIONS.md — cron jobs and maintenance tiers
• CONFIGURATION.md — full config reference