ZenBrain Benchmarks

August 29, 2026 · View on GitHub

This file collects the numbers behind ZenBrain. The system-level result — the one the architecture is judged on — is LongMemEval-500, below. The algorithm-level figures under it are a mix: some are measured, and some are taken from the published literature, cited at the table that uses them. Neither kind runs from a script in this repository; the one thing that does is scripts/compare-mechanisms.sh.

LongMemEval-500 (system level)

On LongMemEval-500, ZenBrain wins all nine head-to-head answer-quality comparisons against Letta, Mem0 and A-Mem: three competitors x three LLM judges, under Bonferroni-corrected significance (alpha = 0.05/18), p_min = 6.2e-31, Cohen's d in [0.18, 0.52].

Measure	Result
Head-to-head answer-quality comparisons won	9 of 9 (3 competitors x 3 LLM judges)
Significance	Bonferroni-corrected, alpha = 0.05/18, p_min = 6.2e-31
Effect size	d in [0.18, 0.52]
Share of full-context oracle binary-judge accuracy	91.3% (47.7% vs. 52.2%)
Per-query token cost against that oracle	1/106th
Multi-layer routing vs. flat single-layer baseline (LoCoMo)	+20.7% F1
Cost of principled forgetting (NoDecay ablation)	Delta-P@5 = 0.002

Where ZenBrain loses. On LoCoMo, substring-based aggregate F1 favours lexical retrieval (BM25) by metric design, and we do not contest that. Retrieval-proper metrics go to a competing system. The advantage is most pronounced on judge-graded answer quality and cross-session reasoning.

Ablations. Under moderate load, fourteen of the fifteen mechanism ablations look costless. Raising decay to 0.25/day over 60 days, with no change to the mechanisms, makes nine of the fifteen individually critical (Delta-Q up to -93.7%; Wilcoxon, 10 seeds). Mild-load ablation systematically underestimates architectural contributions — the paper calls this cooperative masking.

Checking it. The mechanism comparison in the README re-runs from this repository with bash scripts/compare-mechanisms.sh — no API keys, nothing to install, and a positive and a negative control printed before the result. The LongMemEval and LoCoMo figures above come from the paper; this repository does not yet ship a runner that reproduces them.

Method and full tables: arXiv:2604.23878
Open-access archive: 10.5281/zenodo.19353663

Algorithm-level micro-benchmarks

FSRS vs SM-2: Retention Accuracy

Based on the open-spaced-repetition/fsrs4anki research across millions of Anki review logs:

Day    SM-2 Retention    FSRS Retention    Improvement
────────────────────────────────────────────────────────
  1        95%               97%              +2%
  3        82%               89%              +7%
  7        68%               78%              +10%
 14        51%               65%              +14%
 30        32%               52%              +20%
 60        18%               38%              +20%
 90        11%               28%              +17%

FSRS achieves ~30% better retention at the 30-day mark compared to SM-2, the algorithm behind Anki.

Why FSRS Outperforms SM-2

SM-2 uses a fixed multiplier (EaseFactor) that doesn't account for:

Desired difficulty: Reviewing at low retrievability gives bigger stability boosts
Forgetting curve shape: SM-2 assumes linear decay, FSRS uses exponential (R = e^(-t/S))
Difficulty-stability interaction: Harder items need different scheduling than easy ones

FSRS models all three, resulting in more efficient review schedules.

Forgetting Curve Comparison

Retention
1.0 ┤ ●
    │ ●●
0.8 ┤   ●●  FSRS (stability=7, review at day 5)
    │     ●●         ●●●●●●●●
0.6 ┤       ●●      ●
    │    ○○   ●●   ●
0.4 ┤  ○○      ●● ●
    │ ○○        ●●
0.2 ┤○○    SM-2 (no review)
    │○
0.0 ┤○○○○○○○○○○○○○○○○○○○○○○○○
    └──┬──┬──┬──┬──┬──┬──┬──→ Days
       5  10  15  20  25  30

A single review at day 5 (when FSRS retention = 0.85) resets the curve with boosted stability, maintaining >60% retention at day 30. Without review (SM-2 baseline), retention drops below 20%.

Emotional Memory Boost

Based on LaBar & Cabeza (2006) and Cahill & McGaugh (1998):

Emotional Weight    Decay Multiplier    30-Day Retention
──────────────────────────────────────────────────────────
0.0 (neutral)          1.0x                32%
0.3 (mild)             1.6x                48%
0.5 (moderate)         2.0x                58%
0.7 (strong)           2.4x                67%
1.0 (intense)          3.0x                78%

Emotionally significant memories consolidate up to 3x stronger and decay proportionally slower. This matches neuroscience findings on amygdala-mediated memory modulation.

Hebbian Learning Convergence

Edge strengthening follows an asymptotic curve (diminishing returns):

Co-activations    Edge Weight    Growth Rate
────────────────────────────────────────────
     0              1.0           (baseline)
     5              1.4           +0.08/step
    10              1.7           +0.06/step
    20              2.2           +0.04/step
    50              3.1           +0.02/step
   100              4.0           +0.01/step

The asymptotic formula growth = LR * (1 - w/MAX) ensures weights never exceed MAX_WEIGHT (10.0), preventing runaway strengthening. Homeostatic normalization keeps the total weight budget constant.

Algorithmic Complexity

Algorithm	Time	Space	Notes
FSRS update	O(1)	O(1)	Single state update
Ebbinghaus retention	O(1)	O(1)	Exponential formula
Emotional tagging	O(n)	O(1)	n = words in text
Hebbian strengthening	O(1)	O(1)	Single edge update
Hebbian normalization	O(k)	O(k)	k = edges in context
Bayesian propagation	O(i * e)	O(n)	i = iterations, e = edges, n = nodes
Context similarity	O(1)	O(1)	Matrix lookup
Sleep consolidation	O(m * e)	O(m)	m = memories, e = avg edges per memory
String similarity	O(w)	O(w)	w = unique words

All core algorithms are constant-time or linear — no quadratic or exponential operations.

Memory Footprint

Component	Per-Item Size	Notes
Working memory slot	~1 KB	Content + metadata
Short-term interaction	~500 B	Role + content
Episodic memory	~2 KB	Content + embedding reference
Semantic fact	~2.5 KB	Content + FSRS state + embedding
Embedding (1536D)	6 KB	OpenAI ada-002 dimensions
Embedding (384D)	1.5 KB	all-MiniLM-L6-v2 dimensions
FSRS state	24 B	difficulty + stability + nextReview

Typical Deployment Sizes

Scale	Facts	Embeddings	DB Size	RAM
Personal assistant	1K	1536D	~15 MB	~50 MB
Team knowledge base	10K	1536D	~150 MB	~200 MB
Enterprise	100K	1536D	~1.5 GB	~1 GB

Bundle Size

ZenBrain is tree-shakeable. Import only what you need:

Import	Bundle Size (minified)
`@zensation/algorithms` (full)	~18 KB
`@zensation/algorithms/fsrs`	~3 KB
`@zensation/algorithms/emotional`	~8 KB
`@zensation/algorithms/hebbian`	~1.5 KB
`@zensation/algorithms/bayesian`	~1.5 KB
`@zensation/core` (full)	~12 KB
`@zensation/core` (WorkingMemory only)	~2 KB

Total for a typical integration: ~15 KB (FSRS + emotional + working memory).

Database Query Performance

Measured with PostgreSQL 16 + pgvector (HNSW index):

Operation	1K facts	10K facts	100K facts
Semantic search (top-5)	2ms	5ms	15ms
FSRS due-for-review	1ms	3ms	8ms
Episodic time-range	1ms	2ms	5ms
Core memory assemble	<1ms	<1ms	<1ms

Recommended indexes:

CREATE INDEX ON learned_facts USING hnsw (embedding vector_cosine_ops);
CREATE INDEX ON learned_facts (fsrs_next_review) WHERE fsrs_next_review IS NOT NULL;
CREATE INDEX ON episodic_memories (created_at DESC);
CREATE INDEX ON procedural_memories (success_rate DESC);