Introducing akasha-1

The cognitive OS for AI.

#1 on LongMemEval.500/500.Beats Mem0, Zep, and the federated stack — on one engine.

Memory, vector, graph, and analytics — one engine, one API, no glue code.

See the benchmarks$curl -fsSL install.minds.sh | sh
Proven. Not claimed.

The numbers, against everyone else.

Public, third-party benchmarks lead. The head-to-head infrastructure comparison follows. Every caveat stays visible — that's why the wins are believable.

LongMemEval-S · 500 questions

#1 on the public leaderboard

raw score · single-pass real-retrieval · official GPT-4o judge
Akasha-1
0
agentmemory V4
481
PwC Chronos
478
Mastra OM
474
OMEGA
466
ByteRover
461
Hindsight
457
EverMemOS
415
Zep / Graphiti
356
Mem0
335
OpenAI Memory
265

Competitor figures are published raw scores on the same dataset where available. Judge model and scoring convention vary by system; several report task-weighted averages above their raw counts. We show the full field — leading it is the point.

Recall vs Mem0 — hit_any %

AkashaMem0
LoCoMo
97.46%
92.50%
BEAM-1M
94.56%
70.10%
BEAM-10M
84.09%
50.50%

BEAM top-200 hit_any (%), single-pass retrieval. At 10M tokens, where other systems collapse, Akasha holds 84%. Mem0 figures are its published numbers.

Tokens / query vs Mem0 — lower is better

AkashaMem0
LoCoMo16%
5,830
6,956
BEAM-1M17%
5,547
6,719
BEAM-10M18%
5,691
6,914

Mean tokens retrieved per query. 16–18% fewer than Mem0, every compact surface under 7,000 tokens — higher recall on a smaller context.

CogBench · p99 over 10,000 loops

One engine beats the stack you'd assemble

lower is better · read-your-writes 1.000 for all three
Akasha-1
0.00 ms
Postgres + pgvector + AGE + Citus
32.95 ms
Redis + Weaviate + Neo4j + DuckDB
48.85 ms
VectorSIFT1M

Beats hnswlib on QPS and every latency percentile at matched recall (0.9989–0.9991).

Caveat — hnswlib builds its index faster — the win is serving, not build time.

GraphLDBC SNB · bi-temporal

Faster on all 21 LDBC queries vs Kuzu & Neo4j; 54–359× on bi-temporal vs a Neo4j app-layer.

Caveat — Bi-temporal margins are vs an application-layer baseline, not a native engine.

AnalyticsClickBench · freshness

DataFusion-class speed with 44–48× lower fresh-write read latency.

Caveat — On raw ClickBench OLAP, Akasha is at parity — not faster.

Every figure is reproducible — CSVs, audit hashes, and the harness are public.

Reproduce it yourself
One engine, not four

Teams assemble memory from four databases. We replaced them with one.

KV + Vector + Graph + SQL. One async API, cross-modality joins, ANN inside SQL. No glue code, no sync lag, no four bills.

The stack you'd assemble4 services
Rediskey-value
Weaviatevector
Neo4jgraph
DuckDBanalytics
glue codesync lag4 bills4 failure modes
One Akasha engine1 service
KVVectorGraphSQL
  • One async API
  • Cross-modality joins
  • ANN inside SQL
  • No glue code, no sync lag
  • Read-your-writes 1.000

One substrate, five layers — KV up through cognition.

See the L0–L5 architecture
The Cognitive Cycle

Six stages. One loop. Endless improvement.

Perceive
Attend
Decompose
Reason
Learn
Act
Stage 01 · Signals in

Perceive.

Raw experiences — queries, tool results, sensor reads — flow in. Deduplicated, scored for novelty, batched.

Read the spec
Stage 02 · Spotlight

Attend.

The attention filter ranks signals by salience. Only what matters reaches consciousness via the Global Workspace.

Read the spec
Stage 03 · Structure

Decompose.

Triples persist to the knowledge graph. Episodes flow into episodic memory with source attribution. Embeddings index in HNSW.

Read the spec
Stage 04 · Plan

Reason.

HRM hierarchical planning — H-module decomposes the goal, L-module executes. Neurosymbolic blends neural similarity with symbolic paths.

Read the spec
Stage 05 · Reinforce

Learn.

Reward-modulated Hebbian updates strengthen pathways that worked. EWC protects critical weights. Unused connections decay.

Read the spec
Stage 06 · Output

Act.

The selected action emits. ActionFeedbackAdapter routes the result back into Perceive. The loop closes. The agent improves.

Read the spec
cycle://run/8f3alive
Active stage
Perceive142ms
→ 14 raw events · novelty Δ +0.41
Global Workspace · broadcast
Attend38ms
→ top-5 selected · drop rate 64%
Global Workspace · broadcast
Decompose186ms
→ 23 triples · 5 episodes · 12 vectors
Global Workspace · broadcast
Reason418ms
→ 6 subgoals · ACT stopped at depth 4
Global Workspace · broadcast
Learn92ms
→ pathway 06 → 03 · Δw +0.182
Global Workspace · broadcast
Act26ms
→ action dispatched · feedback armed
Global Workspace · broadcast
cycle 8f3a · loop 12reward Δ +0.182
Multi-Modal Memory

Every memory an agent needs, in one place.

Most AI memory products give you one shape: a vector blob. Minds gives you the full neuroscience-grounded hierarchy — episodic, procedural, resource, vault, working, and associative — unified under one async API.

Episodic
Temporal events with emotional valence, importance, and source attribution.
Procedural
Executable workflows. Run them like functions.
Resource
Documents, blobs, and models. Full-text search included.
Vault
AES-256-GCM encrypted secrets, credentials, keys.
Working
Miller's 7±2 capacity. Decay. Variable binding.
Associative
Modern Hopfield networks. Finish a thought from a fragment.
Episodic · live

Agents remember what they did.

Every session captured with goal, outcome, valence, source. The StreamSparkline shows event importance over time. Auditable forgetting handles GDPR delete-requests cryptographically.

11 event types4 valencesMVCCBLAKE3 auditSource monitoring
memory/episodiclast 24h
Reasoning

Plans like a strategist.
Grounded like a lawyer.

The Hierarchical Reasoning Module (HRM) plans abstractly, executes concretely, and decides when to stop. Neurosymbolic fusion grounds every embedding in a symbolic path — hallucinations become a hardware problem you can fix.

HRM · Hierarchical Reasoning Module
H-Module · Abstract Planner
Goal decomposition · subgoal ordering · termination via ACT
slow
Decompose query
Order subgoals
Decide stop
L-Module · Concrete Executor
Tool calls · KG lookups · vector recall · chain-of-thought
fast
query KG
embed
ANN
ground
emit
Adaptive Compute
ACT
Backend
Burn / Rust
Model footprint
≈2.5 GB
Neurosymbolic · live reasoning path
1
Neural87.0%
Pattern matching & embedding similarity
2
Symbolic92.0%
Rule-based knowledge graph traversal
3
Fusion94.0%
Combined neural + symbolic reasoning
Query
> Why did revenue dip in Q3?
14 triples touched·3 microtheories
Bi-Temporal Knowledge Graph

Know what was true, when it was true, and when you learned it.

Every node and edge tracks two timelines — valid time and transaction time. Time-travel any query. Causal-path discovery, counterfactual queries, microtheory scoping, Merkle provenance.

AlicePaperModelFactDatasetAgentClaimObs
t_valid · 2026-05-19
Spiking Neurons

Process time the way brains do.

Nengo-inspired, Rust-native SNN engine. LIF, Izhikevich, STDP, BPTT. Ideal for tick streams, sensor data, and log anomaly detection.

Spike raster · neurons × time
live
00
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
1.4M
spikes/sec
<1ms
step latency
BPTT
training
By the numbers

Sub-millisecond search.
A brain in your binary.

0K
writes / sec
AkashaKV after WAL group commit. 1,673× the original. 30× Redis.
<0µs
P99 latency
Sub-millisecond reads. Sub-10ms vector search at 10K docs.
0M+
rows / sec / node
DataFusion + Arrow Flight analytics throughput.
0
tests in memory crates
Plus thousands more across kv, graph, analytics, reasoning.
What you get

One substrate. Six superpowers.

2024-Q1
n=100
2024-Q3
n=420
2025-Q2
n=1,200
2026-now
n=4,800
Bi-Temporal

Time-travel any query.

Every fact tracks valid time and transaction time. Know what you believed when.

KV
Vector
Graph
SQL
one
Unified

KV + Vector + Graph + SQL. No glue code.

One async API. Cross-modality joins. ANN inside SQL.

EWC protected92%
Replay buffer76%
Quarantined18%
Continual

Train forever. Never forget.

EWC, SI, Generative Replay, and a learning quarantine that rejects poison.

AES-256-GCM · keyed
BLAKE3 audit chain
OIDC + PKCE
Secure

AES-256-GCM. BLAKE3 audit chain. OIDC.

Multi-tenant from the kernel. Differential privacy guardrails. Air-gappable.

write commit1.6ms
vector search · 10K9.4ms
5-hop graph84ms
SQL · 1M rows212ms
Fast

630K writes/sec. Sub-10ms vector.

LSM-tree with WAL group commit, lock-free buffer, TinyLFU cache.

Smart

Self-improves while you sleep.

The cognitive cycle rewards pathways that work. Hebbian + decay.

Governance · Lineage

Every belief, traced to its source.

Knowledge lineage tracks how every fact was derived. Belief-taint tracking ensures untrusted sources cannot launder into Tier-1 recommendations. Auditable forgetting handles GDPR delete requests cryptographically.

governance/lineage · belief #b8f3a4 depths · 7 nodes
Depth 0
memory
interview transcript
5d ago
import
10-Q filing
5d ago
Depth 1
inference
extract entities
5d ago
inference
extract claims
5d ago
Depth 2
merge
deduplicate claims
4d ago
derivation
compute confidence
4d ago
Depth 3
derivation
Q3 revenue ↓ 12%
4d ago
memory
import
inference
merge
derivation
merkle: 0x4a8c…f81e · verified ✓
The Substrate · AkashaKV

The kernel underneath every memory.

Akasha is the cognitive database engine that makes Minds possible. An LSM-tree KV with WAL, MVCC, and lock-free buffers that hits 630K writes/sec — and the substrate every higher-level layer (memory, reasoning, learning) rides on.

akasha · architecture
L5
Interface · HTTP · gRPC · Arrow Flight · MCP
L4
Reasoning · HRM · Neurosymbolic · Cognitive Cycle
L3
Temporal · SNN · SSM/Mamba
L2
Memory · Episodic · Procedural · Working · Vault · Hopfield
L1
Storage · LSM-tree KV · WAL · MVCC · HNSW · Tantivy · DataFusion
L0
Isolation · AES-256-GCM · BLAKE3 audit · 7-action RBAC
Compose your engine

Launch exactly the engine you need.

Toggle a capability on and watch the architecture, the provisioning command, and the price update live. Start with KV; add graph when you need it.

Capabilities3 selected
Live architecture
Graphmodule
Vectormodule
AkashaKVsubstrate
one async API · cross-modality joins · no glue code
terminal
# provisions exactly what you toggled
$ cogs create my-brain --kv --vector --graph --shared
✓ brain ready in ~2s
Estimated
$0

Every capability is free on the shared pool. No feature wall — graph included.

Hot-reloadable

Start with KV. Add graph when you need it. No migration, no downtime, no re-provision — capabilities snap into a running instance.

Shared Akasha

Your data, isolated and encrypted, on a shared regional pool. Instant start, scale-to-zero economics. For prototyping and cost-sensitive workloads.

Dedicated Akasha

Your own isolated instance with predictable performance, resize on demand. Same encryption and isolation as shared — for production and scale.

Data isolation and encryption are identical on shared and dedicated.

Provision your engine
Developer surface

One call. Five languages. Same answer.

Native SDKs for TypeScript, Python, Rust, and Go — plus the cogs CLI and the mind TUI control room. All five wrap the same daemon API.

hybrid recall — vector + graph + episodic
import { Minds } from "@minds-sdk/typescript";

const minds = new Minds({ apiKey: process.env.MINDS_KEY });

// Hybrid retrieval: vector + graph + episodic, one call.
const memories = await minds.recall({
  query: "Why did Q3 revenue dip?",
  modes: ["vector", "graph", "episodic"],
  timepoint: "2026-10-01",
  topK: 5,
});

for (const m of memories) {
  console.log(m.confidence, m.source, m.text);
}
TypeScript
13 modules
Python
sync + async
Rust
native client
Go
production-grade
cogs CLI
multi-instance
mind TUI
tiling control
Live · global mesh
14,832 agent actions reviewed today
alpha · recalled 14 episodes for 'q3 revenue'
beta · committed 412 triples to gnosis://eng
gamma · HRM plan · 6 subgoals · ACT stopped at step 4
delta · pathway 02→05 reinforced · Δreward +0.31
epsilon · 1,204 vectors indexed · HNSW ef=200
zeta · quarantine accepted 18 new facts · 2 rejected
eta · SNN ensemble 'tick-anomaly' spiked at t=8.4s
theta · MHN completion · 87% match · cue 'auth-flow'
alpha · recalled 14 episodes for 'q3 revenue'
beta · committed 412 triples to gnosis://eng
gamma · HRM plan · 6 subgoals · ACT stopped at step 4
delta · pathway 02→05 reinforced · Δreward +0.31
epsilon · 1,204 vectors indexed · HNSW ef=200
zeta · quarantine accepted 18 new facts · 2 rejected
eta · SNN ensemble 'tick-anomaly' spiked at t=8.4s
theta · MHN completion · 87% match · cue 'auth-flow'
iota · auditable-forget · 14 episodes · BLAKE3 verified
kappa · neurosymbolic · 92.4% fusion confidence
lambda · KG time-travel · t_valid 2024-Q1
mu · cognitive cycle 8f3a complete · 902ms
nu · working memory consolidation · 7 → 3 slots
xi · cross-tenant federation · 4 nodes agreed
omicron · ANN routing · 'embedding-similarity' selected
pi · DataFusion · 4.2M rows · 198ms
iota · auditable-forget · 14 episodes · BLAKE3 verified
kappa · neurosymbolic · 92.4% fusion confidence
lambda · KG time-travel · t_valid 2024-Q1
mu · cognitive cycle 8f3a complete · 902ms
nu · working memory consolidation · 7 → 3 slots
xi · cross-tenant federation · 4 nodes agreed
omicron · ANN routing · 'embedding-similarity' selected
pi · DataFusion · 4.2M rows · 198ms
Production-grade

Enterprise-ready the day you adopt it.

Multi-tenant from the kernel. Cryptographic provenance on every fact. OIDC + PKCE auth via Hydra. SOC 2 and ISO 27001 paths in flight.

AES-256-GCM at rest
Per-namespace keys via HKDF-SHA256.
TLS 1.3 in transit
rustls, no OpenSSL.
BLAKE3 audit chain
Tamper-evident, Merkle-anchored.
Multi-tenant isolation
tenant → reservoir → agent → dataspace.
Self-hostable
Run on-prem, in K8s, or air-gapped.
Differential privacy
Guardrails on analytics queries.
Manifesto

Don'tfine-tunetheLLM.Traintheagent.

Domain-specific intelligence belongs in the substrate that persists, not the weights that get replaced every six months. Minds gives your agent a brain it carries forward — across sessions, across model upgrades, across the lifetime of your company.

Pricing

Every capability, every tier.

No feature walls. Graph, vector, and analytics are the same substrate — so they're on from the free tier up. Start on a shared pool, grow into your own instance, compose exactly what you run.

Free · Shared
$0forever
Build a real agent, free.
Full engine — KV · Vector · Graph · Analytics
  • 50,000 memories / month
  • Generous retrieval — no tight call cap
  • 1 shared namespace · scale-to-zero
  • No credit card · never expires
  • Community support
No feature wall. Graph is on from day one.
Start free
Most popular
Starter · Shared
$19/ month
Production-ready shared Akasha.
Full engine — KV · Vector · Graph · Analytics
  • 1,000,000 memories / month
  • Higher throughput, isolated & encrypted
  • Hot-loading included
  • Storage included
  • Email support
The middle tier the incumbents skip.
Start Starter
Dedicated
$99/ month +
Your own instance. Compose what you need.
Pay for the capabilities you turn on
  • Your own isolated instance
  • Modular — KV, Graph, Analytics, HRM, SNN
  • Resize on demand · predictable performance
  • Add capabilities with no migration
  • Priority support
Compose exactly what you run in the engine builder.
Deploy dedicated
Enterprise
Custom
Brain at your scale, your perimeter.
Everything in Dedicated, your way
  • Self-hosted, VPC, or air-gapped
  • SOC 2 / ISO 27001 in flight
  • Bi-temporal provenance · BLAKE3 audit chain
  • Contractual data isolation (DPA)
  • Custom SLA · 99.99%
  • Dedicated solutions architect
Talk to sales
Your memories are yours.

We never train on private data. Aggregated performance metrics help us improve the engine — and if you ever want to contribute more, it's opt-in, always. How we treat your data