Infinite context

Unlimited memory inside a finite context window.

error_outline

Without memory

Every LLM has a context window, and every long-running project eventually exceeds it. Naive memory systems make this worse by stuffing everything they know into the prompt — blowing out cost and drowning the model in irrelevant detail.

check_circle

With Rekall

Rekall’s progressive-disclosure pattern gives agents effectively unlimited memory: recall returns a compact index of what exists, and the agent expands only the entries it actually needs. Months of accumulated memory stay accessible without ever flooding a prompt.

How it works

Recall returns an index

A query returns lightweight summaries — titles, timestamps, relevance — not full payloads. Hundreds of memories fit in a few hundred tokens.

The agent expands selectively

Only the two or three entries that matter get pulled in full. The context window carries signal, not archive.

Consolidation keeps the index sharp

Background workers consolidate, dedupe, and decay old memory so the index stays small and relevant as months of history accumulate.

See it in action

Index first, expand what matters

// Step 1: compact index — cheap, fits any window
const index = await rekall.recallIndex({
  query: 'auth refactor decisions',
})
// → 47 matches as one-line summaries

// Step 2: expand only what the task needs
const full = await rekall.recallFull(index.items[0].id)
// months of memory, a few hundred tokens of prompt