Skip to content

Why I Built emotional-memory: An Affective Field Theory for LLM Agents

Every memory system for language model agents faces the same fundamental question: what makes one memory more likely to surface than another?

Most answers are pragmatic: semantic similarity, recency, importance score. These work well for factual recall. But human memory is not primarily factual — it is affective. We remember what moved us.

emotional-memory is my attempt to build a memory system that takes this seriously.

The problem with semantic-only retrieval

Consider a user who says: "I'm terrified about the presentation tomorrow."

A system that retrieves memories purely by cosine similarity will surface other mentions of "presentation" — regardless of emotional tone. It might return a cheerful memory of a successful talk, a neutral reminder about slide formatting, or a colleague's enthusiastic feedback.

What the user actually needs is mood-congruent retrieval: memories that resonate with their current emotional state — past anxiety, moments of similar vulnerability, or coping strategies from comparable situations.

Mood-congruent memory is not speculative. Gordon Bower demonstrated it empirically in 1981: people are significantly more likely to recall material that was encoded in a matching emotional state. Cognitive systems that ignore this leave real retrieval quality on the table.

Affective Field Theory — five layers

I designed Affective Field Theory (AFT) as a five-layer model, each grounded in established psychology:

Layer Component Basis
1 CoreAffect — valence × arousal Russell 1980 circumplex
2 AffectiveMomentum — velocity + acceleration Spinozist affect dynamics
3 MoodField — slow PAD background Heidegger Stimmung, EMA
4 AppraisalVector — Scherer's 5 SECs Component Process Model
5 ResonanceLink — associative graph Collins & Loftus 1975

These are not independent modules — they interact. A high-arousal memory encodes with a stronger consolidation weight (McGaugh 2004), decaying more slowly. When retrieved, a high prediction-error opens a reconsolidation window (Nader & Schiller 2000) that can update the memory's affective fingerprint. Mood shifts the adaptive weights in retrieval, making emotionally coherent memories more accessible.

What it looks like in code

from emotional_memory import EmotionalMemory, InMemoryStore, CoreAffect
from emotional_memory.embedders import SentenceTransformerEmbedder

em = EmotionalMemory(
    store=InMemoryStore(),
    embedder=SentenceTransformerEmbedder(),
)

# Set the current affective state
em.set_affect(CoreAffect(valence=-0.7, arousal=0.8))  # anxious

# Encode creates a full EmotionalTag snapshot
em.encode("I'm terrified about the presentation tomorrow.")

# Retrieval is filtered by current mood
results = em.retrieve("something difficult is coming up", top_k=3)

The retrieved memories are ranked by a six-signal composite score: semantic similarity, mood congruence, affect proximity, momentum alignment, ACT-R power-law decay, and spreading activation through the resonance graph.

How it differs from existing systems

Systems like Mem0, Letta, and Zep solve the persistence problem brilliantly — they keep large stores searchable across long sessions. What they don't provide is an explicit emotional model. There is no notion of valence or arousal influencing retrieval, no mood state that drifts with the conversation, no reconsolidation triggered by affective prediction error.

emotional-memory is not a replacement for those systems — it is a complementary layer. You could use SQLiteStore as the backend and connect it to an existing LangChain pipeline via EmotionalMemoryChatHistory, adding emotional depth without rearchitecting your stack.

Current state

The library is at v0.5.1:

  • 534 passing tests including 126 parametrized fidelity cases validating 20 psychological phenomena
  • mypy strict — fully typed, zero # type: ignore in source code
  • Extras: [sentence-transformers], [sqlite], [langchain], [viz]
  • MIT license, Python 3.11–3.14

A technical paper is in preparation, with a comparative benchmark against semantic-only baselines and a literature review of 29 related systems.

What's next

  • v0.6.0 — Qdrant/Chroma store adapters, async LangChain support
  • v0.7.0 — OpenTelemetry hooks for production observability
  • Paper — arXiv submission with fidelity validation tables and comparative results

The repository is at github.com/gianlucamazza/emotional-memory. If you build something with it, or find a psychological claim that doesn't hold up, open an issue — I'd genuinely like to know.

Gianluca Mazza — April 2026

Update — May 2026 (v0.9): All three items above shipped as optional extras on the feature/v0.9 branch: QdrantStore ([qdrant]), ChromaStore ([chroma]), and OpenTelemetry spans ([otel]). Roadmap items above reflect the original April expectations and are preserved as a historical record.