The Problem with Static Memory Systems

Traditional AI memory systems require developers to make upfront decisions about how to organize knowledge: What chunk size should I use? Should I use vector search or full-text search? How should I structure my knowledge graph?

These decisions are typically made once and never revisited — even as the nature of the data changes over time. The Adaptive Cognitive Plane (ACP) in CogMem changes this fundamentally.

Memory Strategies as Testable Hypotheses

Adaptive Cognitive Plane architecture

At the heart of the ACP is the concept of a MemoryStrategy — an explicit, versioned hypothesis about how to organize knowledge for a specific scope. A strategy controls:

  • Retrieval policy: FTS-first, vector-first, graph-first, or hybrid
  • Chunking policy: Chunk size, overlap, and splitting strategy
  • Vector policy: Embedding model and similarity threshold
  • Graph policy: Which entity types and relationship types to extract
  • Claim policy: Confidence thresholds and conflict resolution rules

Outcome Signals

Every operation in CogMem emits Outcome Signals — structured feedback about what happened. Did the user find the search result helpful? Was the retrieved context actually used by the LLM? How long did the operation take?

These signals accumulate over time, giving the ACP a rich picture of how well the current strategy is performing.

The Meta-Kernel

The Meta-Kernel is the decision-making component that evaluates strategy effectiveness. It periodically reviews outcome signals and determines whether the current strategy should be adjusted. When it identifies an improvement opportunity, it generates a restructuring proposal with:

  • The specific changes to make
  • The rationale for the change
  • A risk assessment
  • An evaluation window for monitoring

Safety Guarantees

Self-evolving systems need strong safety guarantees. The ACP provides several:

  1. Propose, never execute directly — The Meta-Kernel only proposes changes
  2. Reversible — Every change can be rolled back to the base strategy
  3. Auditable — All decisions emit kernel events with full rationale
  4. Evaluation windows — New strategies are observed before confirmation
  5. Data structures only — The ACP changes data organization, never source code

A Real-World Example

Imagine you're building a research assistant that ingests both academic papers and Slack conversations. Initially, CogMem uses the same chunking strategy for both. Over time, the ACP notices that:

  • Academic papers perform better with larger chunks (preserving context)
  • Slack messages perform better with smaller chunks (each message is a unit)
  • Legal documents benefit from graph-first retrieval (entity relationships matter)
  • Code documentation benefits from FTS-first retrieval (exact terms matter)

The ACP proposes separate strategies for each content type, monitors the results, and confirms the changes when they prove effective.

What's Next

The Adaptive Cognitive Plane is still evolving (pun intended). We're working on more sophisticated evaluation heuristics, cross-tenant strategy sharing, and integration with human feedback loops. Stay tuned for updates.