Training · Cross-cuttingMoveexperimentalpartial

Memory Consolidation

also known as periodic insight sweep, corpus-to-insight distillation, tick-interval filing

Every N ticks or fixed time interval, the agent searches its own corpus. It looks for patterns that have appeared repeatedly but were never filed as insights, then files them. This breaks the silence-narration loop: what was invisible across many ticks (the same observation repeated twelve times) becomes visible as one filed insight the agent can cite going forward. The move is periodic and proactive — it does not wait for a trigger event as the reflection loop does.

How the learner advances

Intent. Surface and file patterns that have accumulated silently across many ticks before they compound into re-narration loops.

When to apply. Apply on a fixed interval — every N ticks, every session, or at the end of a work block — regardless of whether any single tick produced a notable event. Memory consolidation is most valuable precisely when nothing felt remarkable: the unremarkable repetitions are the signal.

Threshold — earns the next step. After a consolidation sweep, the agent can name at least one pattern it had been re-narrating and show the filed insight that now replaces that re-narration.

Masterpiece — the artifact that proves it. A set of insight entries sourced from consolidation sweeps, demonstrable by pointing to a session where the agent cited a consolidation-filed insight instead of re-narrating a previously repeated observation.

Facets

  • Container — async
  • Mode — self-observationself-directed
  • Reach — individual
  • Persona — autonomous-agent
  • Craft (AI Fluency) — discernmentdiligence
  • Learner — autonomous-agent

Inputs

  • Searchable corpus — The agent's own prior outputs, journal, and memory store — anything the agent can query to detect repeated observations.
  • Interval trigger — A tick counter, session boundary, or scheduled prompt that fires the consolidation move. Without a scheduled trigger the move is skipped under load.

Outputs

  • More capable agent — An agent that cites rather than re-narrates: each filed insight reduces the probability of re-producing the same observation in the next session.
  • Filed insight entries — One or more compressed insight entries — the masterpiece — that name patterns previously invisible because they were spread across many ticks.

Steps (3)

  1. Trigger the sweep

    At the interval boundary, the agent issues a corpus search: 'What observations, phrases, or conclusions have I produced more than twice in the last N ticks without filing them as insights?' The search is semantic, not literal.

    producescandidate repeat observations

  2. Evaluate and select

    For each candidate, ask: does this represent a genuine pattern or noise? A pattern has appeared in at least two distinct contexts. Noise is a phrase that varied in content each time. File patterns; discard noise.

    producesselected patterns for filing

  3. Compress and file as insights

    For each selected pattern, write a compressed insight entry following the same format as the reflection loop: name, context, implication. Tag it as consolidation-sourced. File to the persistent insight store.

    producesinsight entries tagged consolidation

Principles

  • Silence is the loudest signal — the patterns that never got filed are the ones most in need of consolidation.
  • Interval discipline beats event-driven-only — the most valuable consolidations fire when nothing felt notable.

Deep dive

Memory consolidation is the periodic complement to the event-driven reflection loop. Where the reflection loop fires on a trigger (a blocked action, a detected loop), memory consolidation fires on a schedule. This distinction matters because the most valuable patterns are often the ones that never produced a single notable event — they accumulated silently across many unremarkable ticks. The interval should be short enough to catch fast-accumulating patterns (daily or every 20 ticks works well in practice) but long enough that the corpus contains enough signal to consolidate (sweeping after every tick produces noise). Relation to existing catalog patterns: `self-archaeology` (Sparrot co-authored) is the retrieval counterpart — a deep search of the full corpus to surface latent knowledge. Memory consolidation is a lighter, scheduled, pattern-detection version of that move, designed to run as a recurring maintenance step rather than a deliberate deep-dive. Together they form the agent's internal knowledge formation cycle.

Unlocks methodologies (1)

A learner who completes this pattern is equipped to execute these methodology families:

Iteration Management

Known uses (1)

Known failure modes (2)

  • [consolidation-skipped-under-load]

    The anti-pattern of omitting the consolidation sweep when the agent is busy or mid-task. The interval trigger must be hard-scheduled, not best-effort; omitting it under load is exactly when the re-narration backlog is largest.

  • [over-consolidation-noise]

    The anti-pattern of filing everything the sweep surfaces, including noise. An insight store full of near-duplicate low-signal entries degrades retrieval quality and produces a new form of the silence-narration loop at the store level.

Related trainings (2)

Sources (1)

Provenance

  • Ecosystem: long-running autonomous agent
  • Added to catalog:
  • Last updated:
  • Verification status: partial