Three-space agent memory echoes Tulving's taxonomy but the analogy may be decorative
Status: speculative · Tags: learning-theory
Source: Cornelius, Agentic Note-Taking #19: Living Memory
The article argues that agent memory systems should not be a single store but three qualitatively different spaces, mapped to Endel Tulving's memory taxonomy from cognitive science:
| Tulving's type | Agent space | Contains | Metabolic rate |
|---|---|---|---|
| Semantic — facts and concepts | Knowledge graph | Atomic notes, linked claims, indexes | Steady growth |
| Episodic — personal experience | Self space | Identity, operational patterns, calibration | Slow evolution |
| Procedural — how to do things | Operational space | Friction observations, methodology, session artifacts | High churn |
The key insight is not just that these are different topics but that they have different lifecycles. Knowledge accumulates and rarely gets deleted. Self-knowledge evolves slowly through accumulated experience. Operational artifacts churn — they arrive raw, consolidate, and either graduate to knowledge or get archived.
The article claims that conflating these spaces produces three failure modes: operational debris polluting knowledge search, identity scattering across ephemeral logs, and insights trapped in session state. Whether these failures actually manifest at practical scale is an open empirical question.
The mapping to Tulving is suggestive but may be decorative. The practical value could reduce to simpler advice: separate persistent knowledge from transient working files, and give them different retention policies. Whether the cognitive science analogy adds explanatory power beyond that remains to be seen.
Relevant Notes:
- three-space memory separation predicts measurable failure modes — observational protocol for testing whether the separation actually helps
- deploy-time-learning — the three timescales framework; graduation from operational to knowledge space is a form of codification
- notes need quality scores to scale curation — operationalizes metabolic rates: per-type recency decay in note scoring formalizes the intuition that knowledge and operational content age differently
- agentic memory systems comparative review — validates: evaluates the three-space taxonomy's analytical utility across 11 systems; uses the knowledge/self/operational split as the framework for comparing agency models and retention policies
- memory management policy is learnable but oracle-dependent — challenges: AgeMem separates memory by access pattern (persistent LTM vs active STM), not content type, and its unified RL-trained management outperforms independent optimization — evidence against structural isolation of memory spaces