Comparison with Existing Systems¶
How emotional_memory positions itself against general-purpose LLM memory
systems. For a synthetic one-line summary see the Home page; for
the underlying numbers see Current Evidence.
| emotional-memory (AFT) | Mem0 | Letta | Zep | LangChain Memory | |
|---|---|---|---|---|---|
| License | MIT | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT |
| Persistence | InMemory / SQLite / Qdrant / Chroma | Qdrant, Chroma, Pinecone, PG, MongoDB | PostgreSQL / SQLite | Neo4j (self-hosted) / Cloud | In-memory / custom |
| BYO embedder | ✅ any Embedder protocol |
✅ (OpenAI default) | ⚠️ partial | ⚠️ partial | ✅ |
| Emotion model | ✅ 5-layer AFT (valence, arousal, dominance, mood, appraisal, resonance) | ❌ | ❌ | ❌ | ❌ |
| Reconsolidation | ✅ APE-gated lability window | ✅ auto update/remove | not measured | ✅ edge invalidation | ❌ |
| Persistent mood state | ✅ MoodField (Heidegger EMA) | ❌ | ❌ | ❌ | ❌ |
| LLM-agnostic | ✅ | ✅ | not measured | ✅ | ✅ |
| LangChain integration | ✅ EmotionalMemoryChatHistory |
✅ official | not measured | ✅ ZepVectorStore | ✅ native |
| Internal fidelity tests | ✅ 127 cases, 20 phenomena (bench-fidelity) | — | — | — | — |
| External benchmark | ✅ LoCoMo (FAIL: F1 0.168 vs 0.271; Pareto Hj1 FAIL) | ✅ LoCoMo, LongMemEval, BEAM | not measured | ✅ DMR, LongMemEval | ❌ |
| Codebase size | ~9k LOC (src/) | >50k LOC | >50k LOC | >50k LOC | ~5k LOC |
Key differentiator: emotional-memory makes affect a first-class, multi-layer part of encoding and retrieval. Compared with the general-purpose memory systems in this table, it emphasizes mood-congruent retrieval, appraisal-conditioned tagging, and APE-gated reconsolidation rather than generic conversational recall alone.
Benchmark caveat: AFT fidelity tests validate psychological invariants (intra-theory). The comparative benchmark in this repo is a controlled synthetic retrieval probe, not a general downstream evaluation of production memory systems. The measured retrieval advantage is regime-specific to affect-discriminative recall under oracle-affect labeling, and on both released third-party emotional retrieval benchmarks cosine is significantly ahead: MADial-Bench (Addendum X, Δ=−0.083 — the corpus rewards counter-congruent supportive recall) and ES-MemEval (Addendum X2, Δ=−0.150 — the QA gold is content-determined and affect-orthogonal, so the affect channel is uninformative noise). See the README section "When NOT to use" for the committed failure regimes.
See also¶
- Current Evidence — the study ladder and claim-to-evidence matrix
- Benchmarks — fidelity, performance, and appraisal-quality suites
- Related Work — full 33-paper survey and comparative positioning