Skip to content

🪞 Metamemory — Self-Reflection

Biological Analog: Metamemory is the awareness of one's own memory processes — "I know I'm forgetting things more often" or "I'm confident I remember this correctly." It's what enables humans to say "I need to write this down" or "Let me double-check that."

The Concept

The introspection system provides analytics and health metrics for the memory system, enabling agents to reason about their own memory state and adapt their behavior accordingly.

Available Insights

flowchart TD
    INTRO["Memory Introspection"] --> TIERS["Tier Distribution<br/><i>counts per tier</i>"]
    INTRO --> TAGS["Tag Distribution<br/><i>top topics by frequency</i>"]
    INTRO --> IMPORTANCE["Importance Stats<br/><i>min, max, mean, stddev</i>"]
    INTRO --> RECALLED["Most Recalled<br/><i>habituation candidates</i>"]
    INTRO --> OLDEST["Oldest Active<br/><i>consolidation candidates</i>"]

    style INTRO fill:#9b59b6,color:white
    style TIERS fill:#4a90d9,color:white
    style TAGS fill:#4a90d9,color:white
    style IMPORTANCE fill:#4a90d9,color:white
    style RECALLED fill:#f39c12,color:white
    style OLDEST fill:#f39c12,color:white
Insight What It Reveals Actionable Signal
Tier distribution Working: 50, Episodic: 12K, Semantic: 200 Imbalance → trigger consolidation
Tag distribution "database": 3K, "error": 2K, "deploy": 500 Topic concentration → diversify or focus
Importance stats mean: 0.45, stddev: 0.12 Low mean → increase surprise sensitivity
Most recalled Memory X recalled 47 times Over-reliance → diversify or suppress
Oldest active Memory Y is 6 months old, never promoted Stale content → consolidate or forget

Adaptive Agent Behavior

An agent can use metamemory signals to self-optimize:

flowchart TD
    CHECK["Agent inspects memory health"] --> C1{"High episodic,<br/>low semantic?"}
    C1 -->|"Yes"| A1["Trigger consolidation<br/><i>'I should sleep on this'</i>"]

    CHECK --> C2{"One memory recalled<br/>47 times?"}
    C2 -->|"Yes"| A2["Diversify recall<br/><i>'I'm over-relying on this'</i>"]

    CHECK --> C3{"Low importance<br/>average?"}
    C3 -->|"Yes"| A3["Increase surprise sensitivity<br/><i>'Everything feels routine'</i>"]

    style A1 fill:#9b59b6,color:white
    style A2 fill:#f39c12,color:white
    style A3 fill:#e74c3c,color:white

Example Scenarios

Observation Agent Response
90% episodic, 2% semantic "I should consolidate — trigger a reflect cycle"
One memory dominates all queries "I'm over-relying on this — increase habituation rate"
Average importance is 0.2 "Most memories are routine — lower the surprise threshold"
500 memories tagged "error" "I'm seeing too many errors — escalate to the user"

Next Steps