Memory · Working

The scratchpad that decays.

Working memory holds the slots an agent is currently using — bounded by Miller's 7±2 capacity model, decaying on configurable curves, with variable binding and attention gating. The bridge between perception and reasoning.

Start free Read the docs
Capabilities

What it does.

Every capability is production-grade. No flags, no betas — these ship the day you adopt Minds.

Miller 7±2
Slot count tunable per agent; default 7. Eviction policy: oldest-low-importance first.
4 decay functions
Linear · exponential · power-law · ACT-R. Per-slot, per-agent.
Variable binding
Slots can reference other slots; binding survives consolidation.
Attention gating
The attention filter chooses which slot the cognitive cycle reads from next.
MVCC
Concurrent reads/writes safe across the cognitive cycle.
Consolidation
Slots that persist long enough are promoted to episodic memory.
API

A few lines is all it takes.

The SDK reflects the conceptual model directly. No glue code, no orchestrator, no learning curve beyond the data shape.

exampletypescript
// Put something in working memory.
const slot = await minds.memory.working.put({
  key: "current_question",
  value: "Why did Q3 revenue dip?",
  importance: 0.9,
  decay: { fn: "act-r", rate: 0.18 },
});

// Bind another slot to it.
await minds.memory.working.bind({
  key: "current_evidence",
  to: slot.id,
  value: { sources: ["10-Q", "CFO interview"] },
});

const snapshot = await minds.memory.working.read();
// snapshot.slots is bounded by capacity, sorted by importance × recency
Specs

What it costs, what it guarantees.

Performance
Default capacity
7 slots
Min/max capacity
3 · 11
Decay functions
linear · exponential · power-law · ACT-R
Slot read latency
<0.3ms
Tests
155 (working-memory crate)
Guarantees
Durability
fsync per WAL commit
Isolation
MVCC · snapshot
Audit
BLAKE3 Merkle chain
Encryption
AES-256-GCM per-namespace
Concurrency
100K+ ops/sec
Keep exploring
Episodic memory
Working slots consolidate here.
Cognitive Cycle
How attention reads from working memory.
Associative memory
Pattern completion from partial cues.

Build it on Minds.

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

Start free View pricing