Architecture Overview

The problem with static memory

Most AI memory systems work like a filing cabinet: store a document, retrieve it later by similarity. This approach has two fundamental problems:

1. No forgetting. Every memory has equal weight regardless of age. A user’s address from three years ago competes equally with their current address.
2. No context. Memories exist in isolation. There’s no way for retrieving one fact to activate related facts.

Human memory solves both problems. It decays over time (forcing prioritization), strengthens with use (surfacing important information), and forms associative networks (enabling multi-hop reasoning).

Cognitive Memory’s architecture

The system implements six biologically-inspired mechanisms:

1. Ebbinghaus decay with floors

Every memory has a retention score R(m) that decays exponentially over time. The decay rate depends on the memory’s type, stability, and importance. Core memories have a retention floor of 0.60 — they dim but never vanish.

See Decay model and Decay floors.

2. Memory categorization

Memories are classified into four categories, each with different decay characteristics:

Category	Base decay (days)	Floor	Description
Episodic	45	0.02	Events with time/place context
Semantic	120	0.02	Facts, preferences, relationships
Procedural	infinite	—	Skills, routines (never decay)
Core	120	0.60	Identity-defining information

See Memory types.

3. Retrieval-weighted scoring

Search results are scored by sim(m,q) * R(m)^0.3 — cosine similarity modulated by retention. The exponent 0.3 softens the decay penalty so faded memories can still surface when highly relevant.

See Retrieval scoring.

4. Two-tier retrieval boosting

When a memory is retrieved, its stability increases. Direct matches get a larger boost (0.1) than associatively-activated memories (0.03). Both are modulated by a spaced repetition factor: longer gaps between retrievals produce bigger boosts.

See Retrieval boosting.

5. Associative memory graph

Memories form bidirectional links through two mechanisms:

• Synaptic tagging at ingestion: memories encoded in the same session with semantic overlap are linked.
• Co-retrieval strengthening: memories retrieved together have their link weights increased.

Links decay exponentially with time constant 90 days.

See Associations.

6. Tiered storage lifecycle

Memories move through three tiers:

• Hot: full embeddings, vector-searchable
• Cold: ID-accessible only, no vector search
• Stub: archived summary, minimal footprint

See Tiered storage.

How it fits together

The full pipeline from ingestion to retrieval:

Conversation text
      |
      v
  LLM Extraction (narrator prompt)
      |
      v
  Memory objects with category + importance
      |
      v
  Embedding + conflict detection
      |
      v
  Synaptic tagging (ingestion-time associations)
      |
      v
  Storage in hot tier
      |
      |--- Search query arrives --->
      |
      v
  Embedding similarity search (hot store)
      |
      v
  R^alpha temporal scoring
      |
      v
  Associative expansion (linked memories)
      |
      v
  Direct + associative boosting
      |
      v
  Core promotion check
      |
      v
  Sorted results returned

Periodic maintenance (tick()) handles cold migration, TTL expiry, and consolidation in the background.

See Scoring pipeline for the complete retrieval flow, and Ingestion for the extraction pipeline.