Summary

Add trigger metadata to agent sessions (values: user, cron, subagent, hook) and exclude trigger: cron sessions from memory-core Dreaming ingestion by default, with opt-in override per cron job.

Problem to solve

memory-core's Dreaming pipeline ingests cron-triggered sessions (heartbeat pings, scheduled automation, email-watch jobs) into __dreaming_sessions__ identically to user-initiated conversations. On any deployment that uses openclaw cron non-trivially, this starves the REM and deep-promotion phases of real signal.

Concrete example from a fresh memory-core install, day 1, single-agent personal Donnie on Discord (OpenClaw 2026.4.14):

REM themes (confidence 1.00):
  - heartbeat-ok    60 memories
  - user            45
  - assistant       44
  - heartbeat.md    30
  - home/hyamie     30
  - openclaw/workspace/heartbeat.md  30
  - read            30

Deep phase: zero promotions. "No strong candidate truths surfaced."

The corpus file memory/.dreams/session-corpus/2026-04-13.txt contains 60 lines of:

User: Read HEARTBEAT.md if it exists [...] reply HEARTBEAT_OK.
Assistant: HEARTBEAT_OK

Actual signal (family context, golf communications, project work) scores identically at confidence 0.58 per line, and gets buried. The deep phase's promotion gates (minScore 0.8 / minRecallCount 3 / minUniqueQueries 3) can never fire on real content because it's outweighed by cron repetition.

This is closely related to #63921 (transcript wrapper noise), but that issue's proposed fixes (stopword filter, wrapper stripping, role-token normalization) operate at the token level and don't address cron-content pollution. Heartbeat content is valid English prose that repeats — no stopword list can remove it without also removing legitimate user content.

Related: #63921 #59095 #29442.

Proposed solution

1. Session-level trigger metadata (new)

Every session gets a trigger field set at creation, written to session state and available to plugins:

The [cron:<id> <name>] preamble already emitted by openclaw cron is sufficient signal — formalize it as session metadata at session open time rather than as in-band text.

2. memory-core Dreaming opt-out for non-user triggers (new default)

memory-core config gains a dreaming.ingest.triggers array. Default:

{
  "memory-core": {
    "dreaming": {
      "ingest": {
        "triggers": ["user"]
      }
    }
  }
}

Non-matching sessions are skipped during __dreaming_sessions__ enumeration. Operators who want a cron workload reflected on explicitly can set ["user", "cron"].

3. Per-cron-job override (fine-grained)

openclaw cron job definition gains dreamingIngest: boolean (default false if not set, same outcome as trigger filter). Lets someone whitelist a specific high-signal cron (e.g. a daily journaling prompt) without opening the floodgates.

4. Backward compatibility

Existing sessions without a trigger field default to user — matches current behavior, zero regression for users who don't run cron jobs.

Alternatives considered

• Token/phrase blocklist only (what [Bug]: Dreaming/REM surfaces raw session-corpus metadata (untrusted, user, assistant) instead of meaningful reflections #63921 proposes): insufficient for cron content because heartbeat messages are valid English, not structural boilerplate. Scales poorly — every new cron job needs manual phrase additions.

• Entropy/novelty gating (SimpleMem arxiv 2601.02553, entity novelty + semantic divergence threshold τ=0.35): principled but substantially more implementation work, requires embedding every exchange pre-ingestion, and gives fuzzy results when the right answer is a hard "skip." Good candidate for a Phase 2 layer on top of trigger filtering.

• Min-token exchange gate: simple, catches heartbeat (4-6 tokens) but not longer cron runs like email-watch playbook invocations. Useful belt-and-suspenders but not sufficient alone.

• Per-session importance weighting pre-TF-IDF (LangMem pattern — multiply phrase frequencies by a session weight before REM theme extraction): equivalent outcome to trigger filtering but more invasive change to the theme extractor. Trigger-based ingestion gating is simpler and affects fewer components.

• Do nothing, rely on volume: wait for real user conversation to eventually outweigh cron noise. Doesn't work — cron runs on a fixed cadence, so the ratio stays roughly constant as real usage grows.

Impact

Affected: Anyone using openclaw cron with memory-core Dreaming enabled. As memory-core becomes the default memory plugin and cron gets used for automation (heartbeats, scheduled email checks, daily journaling prompts, periodic agent-driven tasks), this affects a growing share of users.

Severity: High for users who rely on Dreaming-promoted long-term memory. On my setup, deep phase produced zero promotions in week 1 despite real conversational content being present, because cron sessions dominated the corpus. Effectively breaks Dreaming's value proposition on any non-trivial cron deployment.

Frequency: Every Dreaming run. Daily for most users (default cron 0 3 * * *).

Consequence: Promoted "lasting truths" become noise or nothing. Operators lose trust in memory-core's consolidation layer and either disable Dreaming, disable cron automation, or manually prune session corpus files — all of which defeat the point.

Evidence/examples

Prior art:

• Mem0 issue What we found after auditing 10,134 mem0 entries: 97.8% were junk mem0ai/mem0#4573 — "97.8% were junk" audit, same shape of problem. Their filtering.ts (PR feat(openclaw): improve extraction quality with noise filtering, deduplication, and better instructions mem0ai/mem0#4302) is a message-level blocklist; author concluded it's insufficient alone and that source-level gating is the real fix.
• Redis agent-memory-server (github.com/redis/agent-memory-server) — uses source_type / session_id / agent_id metadata at write time, consolidation pipeline filters by source_type.
• Letta/MemGPT — memory architecture separates system-originated from user-originated content at the block level before archival.
• LangMem — per-session importance weights applied before theme extraction.
• SimpleMem (arxiv 2601.02553) — entity-novelty + semantic-divergence threshold gate, cites "phatic chit-chat and redundant confirmations" as the canonical discard case.

Local evidence files (available on request):

• memory/.dreams/session-corpus/2026-04-{13,14,15}.txt — 89 lines total, 60+ from heartbeat cron
• memory/.dreams/phase-signals.json — shows every heartbeat line counted equally with every real-conversation line
• memory/2026-04-15.md — REM output above, Axis B themes dominating

Additional information

Willing to contribute a draft PR if maintainers confirm direction. Two candidate scopes:

• Minimal (recommended first): add trigger metadata field to session creation, add dreaming.ingest.triggers config to memory-core, filter at ingestion.
• Full: also add dreamingIngest: bool to openclaw cron job schema and admin UI.

Cross-link: #63921 covers the orthogonal Axis A (transcript wrapper noise). Adding a comment there with this setup's repro data to confirm regression persists in 2026.4.14.

Happy to also seed a phrase blocklist for the cron preambles (HEARTBEAT_OK, HEARTBEAT.md, NO_REPLY, [cron:...]) as a lower-cost Phase 0 fix while trigger metadata lands — these are safe to stopword because they're literal protocol tokens, not semantic content.