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 — six memory types, unified under one async API, with auditable forgetting and cryptographic lineage.

Start free Read the spec
The six memory types

One async API. Six memory shapes. Zero glue code.

Each memory type is a first-class primitive with its own access patterns, lifecycle, and audit trail — and they all share the same substrate, so cross-modal retrieval is a single call.

Episodic
Temporal events with emotional valence, importance scoring, and source attribution. 11 event types, 4 valences, full provenance.
Procedural
Executable workflows. Store agent routines as functions that can be run, composed, and improved over time.
Resource
Documents, blobs, and models with full-text search. Tantivy-backed indexing across every artifact your agent touches.
Vault
AES-256-GCM encrypted store for credentials, tokens, and keys. Four sensitivity tiers, key rotation, zeroize on drop.
Working
Miller's 7±2 capacity model. Linear/exponential/power-law/ACT-R decay. Variable binding and attention gating.
Associative
Modern Hopfield networks. Finish a thought from a fragment. 30% noise tolerance, exponential capacity.
Hybrid retrieval

Vector, graph, and episodic — one call.

The MetaMemoryManager fuses results across memory types in a single async call. No more wiring three databases together and praying the responses align.

hybrid recall — fuse all memory modestypescript
import { Minds } from "@minds-sdk/typescript";

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

// One call. Fuses across vector, graph, and episodic memory.
const memories = await minds.recall({
  query: "Why did Q3 revenue dip?",
  modes: ["vector", "graph", "episodic"],
  timepoint: "2026-10-01",         // bi-temporal — query the past
  microtheory: "finance",          // scope to a worldview
  source: ["observed", "inferred"],// filter by attribution
  topK: 5,
});

for (const m of memories) {
  console.log(m.confidence, m.source, m.text);
}
Auditable forgetting

GDPR delete-requests, cryptographically.

Every deletion is justified, signed, and tamper-evident. Knowledge laundering is structurally impossible. Belief-taint tracking ensures untrusted sources can never reach Tier-1 recommendations.

Memory architecture
Memory types
6 (episodic, procedural, resource, vault, working, associative)
Capacity model
Miller 7±2 (working)
Decay functions
linear · exponential · power-law · ACT-R
Source attribution
observed · inferred · imagined · reported · reconstructed
Bi-temporal
valid time + transaction time
Concurrency
MVCC, 100K+ ops/sec
Tests
451+ across memory crates
Compliance & audit
Audit chain
BLAKE3, Merkle-anchored
Encryption at rest
AES-256-GCM, per-namespace keys
Key derivation
HKDF-SHA256
Forgetting
auditable · justified · signed
Lineage
every belief traced to its source
Quarantine
verification gates before promotion
GDPR
cryptographic delete-receipts
Keep exploring
Reasoning
HRM hierarchical planning + neurosymbolic fusion.
Knowledge Graph
Bi-temporal property graph with provenance.
Cognitive Cycle
Perceive → Attend → Reason → Decompose → Learn → Act.

Build a mind that remembers.

Start with the SDK. Ship in an afternoon. Run anywhere — Akasha Cloud, on-prem, or air-gapped.

Start free View pricing