causal-memory

Benchmarks

All results reproducible from the benches/ directory. Protocol details in docs/benchmarks.

CausalEval — the causal memory benchmark

Typed DAGs are generated deterministically; conversations are narrated from the graph; gold answers are derived from graph structure — zero hand annotation, zero ambiguity.

Capabilitycausal-memorymem0What it tests
C7 Update100%80%Supersede old belief after falsification (soft superseded_by annotation)
C3 Counterfactual95%80%Choosing between alternatives with known outcomes
C2 Intervention75%40%Forward prediction: "if X again, what happens?"
C4 Inhibition80%50%Root-cause fix vs blast-radius limiter (prevented edges)
C1 Attribution85%90%Backward causal chain → root cause
C5 Temporal-causal90%90%Ordering on a causal chain
C6 Lesson transfer20%30%Cross-task analogy via meta edges (open limitation)
Overall78%65%CausalEval v13 · 140 questions, 20 graphs, same LLM, same judge

Fact-recall benchmarks

On traditional fact-recall suites causal-memory is competitive but does not beat mem0 — fact recall is mem0's specialty, not where causal-memory adds value.

Benchmarkcausal-memorymem0Note
LoCoMo (strict judge)79.1%91.6%mem0's home turf
LongMemEval-S76.4% @ 11.5K tok/q94.4% (official)single-model stack vs platform stack
Memora MPA67.4%71.8%−4.4pp
Compaction survival100%45%External table = immune to compaction
Agent repeat-mistake33%67%−34pp on trap-world

Capability tests

322 workspace tests covering capabilities no fact store (mem0, Zep, Letta) offers.

CapabilityWhat it proves
Prevented-edge warningprevented edge spreads −0.3 activation (GABA analogue)
Trace-cause attributionBackward CSR traversal finds root cause
Multi-hop causal chainForward K-hop spreading reaches 2–3 hop outcomes
Inhibitory filteringPrevented outcomes appear as negative, not false positives
Intervention comparisonSame outcome: +0.9 for "skip tests", −0.3 for "add tests"
SWR consolidationLTP strengthens replayed edges, LTD weakens unvisited, GC forgets dormant
Q-value dynamicsGood decisions rank higher; Bellman propagates to parents
Novelty entropyDiverse experience triggers consolidation; uniform does not
Meta-edge miningCross-session pattern discovery (similar_to / repeated)
Hebbian co-occurrenceRepeated co-activation strengthens connection