[YOUR VOICE] The Claim
Shared agent memory is easy to build and hard to trust. The challenge isn’t storage or retrieval — it’s knowing which memories are still valid, which have drifted, and which were wrong to begin with. Memory without metacognition is just a pile of stale assertions.
The Mechanism
The Metacognitive Memory System serves 11+ agent entities across 44+ projects. Three mechanisms prevent memory decay:
- Drift detection — identifies when stored memories conflict with current project state
- Extraction normalization — ensures memories from different agents and sessions use consistent formats and terminology
- Metacognitive reliability scoring — each memory carries a confidence assessment that degrades over time and gets reinforced by corroborating evidence
MISSING — Architecture diagram showing the three-layer memory system (OpenMemory for session memory, Basic Memory for file-backed context, project _docs/ for ground truth)
MISSING — Specific examples of drift detection catching stale memories
The Evidence
MISSING — Memory system metrics: number of memories stored, drift detection rate, false positive rate on reliability scores
MISSING — Before/after: agent task quality with and without metacognitive memory
[YOUR VOICE] Implications
MISSING — Why every multi-agent system needs metacognition about its own memory. The cost of trusting stale context.
Open Questions
- How should memory reliability scores decay over time?
- Can agents self-assess the quality of their own memory writes?
- What’s the minimum memory infrastructure needed for a 3-agent vs. 11-agent system?
Reference Documents
| Document | What it covers |
|---|---|
| Memory-System _docs/ | MISSING — Architecture and implementation |
| Interagent memory patterns | MISSING — Cross-agent memory sharing protocol |
| Operator runbook | MISSING — Getting started guide for the memory system |