Skip to content

ACT-R Adaptations for TeaParty

How TeaParty's proxy memory system adapts ACT-R declarative memory for agent gate decisions. This document covers only where we depart from vanilla ACT-R. For the theory, equations, and standard parameter values, see research/act-r.md.

Related: - memory.md — motivation and migration plan - mapping.md — chunks, traces, retrieval implementation - sensorium.md — two-pass prediction and learned attention - ../index.md — conceptual design

Interactions, Not Seconds

In ACT-R, t in the base-level activation equation is measured in seconds. We measure t in interactions: gate decisions, dialog turns, observations. Each interaction advances the clock by 1.

Between sessions, nothing happens. If the agent handles 5 decisions on Monday and none until Thursday, wall-clock decay would erode Monday's memories over 3 idle days. Interaction-based decay does not advance. No interactions means no decay, which is correct because nothing happened to make the memories less relevant.

Anderson & Schooler's empirical basis is event-based. Their 1991 analysis measured word occurrences in newspaper headlines, child-directed speech, and email. The power-law pattern they found was in events (how many headlines ago did this word last appear?), not in seconds. The environment's statistical structure is event-based. So should the memory system's. (Note: their primary unit of analysis was days for NYT/email and utterance intervals for speech. The "event-based" framing is a reasonable interpretation of their methodology, not a direct quote from the paper.)

Experience scales with activity, not calendar time. An agent that handled 100 interactions over a busy week has far more trace accumulation than one that handled 5 over the same calendar period. The memory should reflect experience, not elapsed time.

Parameter Choices

We depart from ACT-R defaults for noise and retrieval threshold.

Parameter ACT-R Default Our Value Rationale
Decay (d) 0.5 0.5 No departure. Empirically validated.
Noise (s) NIL (disabled) 0.08 Calibrated so noise std dev (~0.145) stays below typical signal differences (~0.3). Perturbs ranking without dominating it. Validated via noise-scale sensitivity analysis in the ablation harness.
Retrieval threshold (tau) 0 -0.5 Admits chunks with slightly negative activation. Desirable in a low-interaction system (50-200 lifetime interactions) where useful chunks may hover near zero. Needs empirical calibration.

Caveat on d = 0.5. Anderson & Schooler's corpora are high-volume streams (thousands of observations). Proxy gate interactions are sparse — perhaps 50-200 total lifetime interactions. The power-law form is the right functional shape, and d = 0.5 produces moderate decay. But whether 0.5 is optimal for this interaction regime is an open empirical question.

Context Sensitivity via Embeddings

ACT-R uses symbolic spreading activation for context sensitivity: activation spreads from the current goal through associative links between chunks. The spread depends on fan (how many chunks share an association) and source activation.

We replace this with vector embeddings and cosine similarity. Each chunk has up to 5 independent embedding dimensions (situation, artifact, stimulus, response, salience). At retrieval, cosine similarity between query embeddings and chunk embeddings provides the context-sensitivity signal.

This is a fundamental departure. ACT-R's spreading activation is symbolic, discrete, and structure-aware. Cosine similarity over embeddings is continuous, learned, and structure-blind. The trade-off: we lose the ability to reason about structural relationships between chunks, but we gain the ability to match on semantic meaning without manually defining associative links.

Two-Stage Retrieval

ACT-R retrieves the single chunk with highest activation above threshold. We retrieve a ranked set using a two-stage process:

  1. Activation filter: all chunks with B > tau survive (standard ACT-R)
  2. Composite ranking: survivors are scored by activation_weight * normalized_B + semantic_weight * cosine_sim + noise and the top-k are returned

The composite score mixes ACT-R activation with semantic similarity. This is not ACT-R — it is a hybrid that uses ACT-R's activation as one signal among two. The design doc for this is mapping.md.

Normalization is min-max over the candidate set. This makes the effective weight of activation relative to the current query, not absolute.

Reinforcement Model

In ACT-R, a chunk is reinforced (gets a new trace) every time it participates in a production rule firing. In our system, reinforcement happens at two points:

  1. On chunk creation — the initial trace
  2. On explicit reinforcement — after the proxy has consumed retrieved memories and produced a response, the caller explicitly reinforces the chunks that were used via reinforce_retrieved()

Retrieval itself does not add traces. This departs from some ACT-R implementations where retrieval automatically reinforces, but aligns with the ACT-R specification that reinforcement occurs when a chunk is "specifically retrieved for a task and actively referenced by a production rule."

What We Don't Use

Several ACT-R mechanisms are not implemented:

  • Spreading activation — replaced by embedding similarity (see above)
  • Partial matching — ACT-R can retrieve chunks that partially match a query, with a penalty. We use cosine similarity instead, which provides a continuous match score.
  • Production compilation — ACT-R's procedural learning mechanism (combining two production rules into one). Not applicable to our system.
  • Utility learning — ACT-R learns which production rules to prefer via reward. Not applicable; we have a separate confidence calibration mechanism.
  • Soft-threshold retrieval probability — the logistic P(retrieve) function from ACT-R eq. 4.4. We use a hard threshold (B > tau) for stage 1 filtering. The soft threshold could replace this; it's documented in research/act-r.md for future consideration.