Skip to content

CoALA Memory Taxonomy as a Classifier for Harness Artifacts

Use CoALA's four memory types — working, episodic, semantic, procedural — to classify harness artifacts and surface capability gaps.

CoALA is a descriptive framework that maps language agents onto four memory types inherited from Soar and ACT-R (Sumers et al., 2024). Applied to a Claude Code, Copilot, or Cursor harness, it names what each artifact is and which memory type has no representative — not how to implement any of them.

When the classifier earns its keep

The taxonomy pays off when a harness has enough memory surfaces that absence is non-obvious. Apply it when:

  • The harness has at least three distinct persistence surfaces — CLAUDE.md, a transcript store, a RAG index — and you want to know what is missing.
  • You are auditing a harness you did not build. Naming each artifact in CoALA terms forces a complete inventory; ad-hoc inspection skips categories that are not there.
  • A capability gap is observable but not localized. "The agent keeps making the same mistake across sessions" maps to "no episodic memory or no learning action into procedural memory" once the inventory is in CoALA terms (CoALA §3).

Skip it when the harness is single-session, single-file, and single-developer — there is nothing to classify.

The four types, defined by the source

Definitions are quoted from CoALA §3; paraphrasing them loses the framework's grounding.

Type CoALA definition (Sumers et al., 2024)
Working "Maintains active and readily available information as symbolic variables for the current decision cycle."
Episodic "Stores experience from earlier decision cycles … history event flows, game trajectories from previous episodes, or other representations of the agent's experiences."
Semantic "Stores an agent's knowledge about the world and itself" — readable from external sources or writeable from reasoning.
Procedural "Two forms … implicit knowledge stored in the LLM weights, and explicit knowledge written in the agent's code."

Working memory is transient and within-cycle. Episodic, semantic, and procedural are long-term (§3.1.1–3.1.4).

The mapping table

The classifier resolves cleanly onto coding-agent artifacts. The mapping is the diagnostic instrument.

CoALA type Concrete harness artifact What it's good at What it's bad at
Working The live context window; in-turn scratchpads Holding the current goal, recent tool results, and intermediate reasoning Persistence — evaporates at session end and degrades inside the context-window dumb zone
Episodic Session transcripts; session recap files; execution traces; memory synthesis from execution logs Recalling what was tried in similar past situations Generalising across episodes — raw transcripts re-injected verbatim crowd context without synthesis
Semantic RAG indexes, vector stores, knowledge graphs, codebase symbol indexes, structured domain retrieval Looking up facts — API signatures, schema, file locations Tracking whether the fact is current; mixing facts and experiences under uniform decay produces a category error (Roynard, 2026)
Procedural CLAUDE.md / AGENTS.md, skills, slash commands, hooks, the agent's own code Encoding how to act — the procedure the agent runs every time Reflecting recent experience — procedural memory is rarely updated mid-session; learning is offline by design (CoALA §3.1.4)

The CoALA paper applies the same classifier to research systems: ReAct has working memory only; Voyager has procedural learning; Generative Agents combine episodic and semantic (Table 2). The mapping above is the equivalent for production coding-agent harnesses.

Using the taxonomy as a diagnostic

The diagnostic is a single question per type: what artifact, if any, plays this role in our harness? Each blank answer is a candidate capability gap.

graph TD
    A[Inventory harness artifacts] --> B[Working: context window]
    A --> C[Episodic: transcripts / recap?]
    A --> D[Semantic: RAG / KG?]
    A --> E[Procedural: CLAUDE.md / skills?]
    B --> F[Any blank?]
    C --> F
    D --> F
    E --> F
    F -->|Yes| G[Localised capability gap]
    F -->|No| H[Audit operational quality<br>not type coverage]

    style G fill:#5a2d2d,stroke:#4a4a4a,color:#e0e0e0
    style H fill:#2d5a2d,stroke:#4a4a4a,color:#e0e0e0

A harness with only working and procedural memory cannot learn from past sessions — the absence of episodic memory predicts the failure mode without testing. A harness with episodic but no semantic memory will repeatedly re-derive facts that should live in a lookup, the consolidation gap tiered memory architecture closes. The classifier names these absences directly.

Why it works

The four-type split is not arbitrary. CoALA inherits it from Soar and ACT-R — cognitive architectures whose purpose is to expose which capabilities a memory system has and lacks (Sumers et al., 2024 §2.3) — and re-applies that mapping to language agents. Sumers et al. demonstrate the mechanism in Table 2: noting ReAct "lacks semantic or episodic memory" immediately predicts its inability to retrieve from or learn across episodes (§3). The classifier earns its keep when absence is the diagnostic signal — what the cognitive-architecture inheritance is designed to make legible.

When this backfires

The classifier describes content type, not operational mechanics. Treating the label as a build spec produces real architectural bugs.

  • Category error: facts versus experiences under uniform decay. Labeling a writeable RAG index "semantic memory" implies it inherits properties — abstraction, consolidation, integration with episodic memory — that it lacks. Roynard (2026) identifies this directly: CoALA "lacks an explicit Knowledge layer with its own persistence semantics," so harnesses applying the same update mechanics to facts and experiences silently corrupt one or both.
  • Ambiguous boundaries on writeable stores. An index the agent queries and writes to is partly semantic memory, partly an external action target. The classifier produces a category dispute rather than insight; the memory/environment boundary is not crisp for digital agents as it is for embodied ones.
  • Trivial agents pay ceremony without yield. A one-file CLAUDE.md and a context window need no taxonomy. The classifier's value is proportional to how much memory surface there is to classify; on small surfaces the description is the inventory.
  • Mislabeling implies false architectural depth. A team claiming "all four CoALA memory types" can ship a harness whose semantic memory is one embedded JSON file and whose episodic memory is the session window — labels present, capabilities absent. Type coverage is necessary but not sufficient; operational quality is a separate audit.

Lead with the artifacts, then use CoALA's full three-axis framework to name them and surface gaps. Do not let the labels imply mechanics the underlying artifacts do not have.

Example

Inventorying a typical Claude Code coding-agent harness with CoALA:

Artifact in this harness CoALA type Notes
Live context window Working Standard; subject to dumb-zone degradation past ~50% fill
.claude/state/*.jsonl session logs, session recap at compaction Episodic Recap is goal-shaped, not raw transcript — already a step beyond verbatim replay
No vector index, no knowledge graph Semantic — missing Reader looks up library APIs via Bash + WebFetch every time; gap predicts repeated re-derivation cost
CLAUDE.md + .claude/skills/ + .claude/agents/ + hooks Procedural Rich procedural layer; the harness's strongest type

The inventory says: missing semantic memory. The capability prediction follows directly: the agent will keep paying retrieval cost on facts that should sit in a lookup, and any "agent kept getting the API signature wrong" failure traces to that absence — not to a working-memory or procedural-memory bug. Adding a small semantic index (context hub, repository map pattern) is the next architectural move.

Key Takeaways

  • CoALA's four memory types map cleanly onto production coding-agent artifacts; the mapping is the diagnostic tool.
  • The classifier earns its keep on multi-layer harnesses where absence of a memory type is non-obvious by inspection; it adds ceremony without value on trivial agents.
  • The four-type lens describes content type, not operational mechanics — a labelled type can still be operationally broken, and writeable stores blur memory-versus-environment boundaries.
  • Roynard (2026) identifies a category-error risk: applying uniform decay or update mechanics to facts and experiences under the same "long-term memory" label silently corrupts one or both.
  • Use the taxonomy to name what you already have, not as a build recipe — the implementation pages remain the source of truth for how each memory type actually works.
Emerging — single recent source Last reviewed: 2026-06-12
Feedback