Infinite Context

Rekall's progressive disclosure pattern that lets agents access unlimited memory without exhausting their context window.

LLM context windows are finite, but memory stores can be massive. A coding agent might accumulate thousands of memories across sessions — project decisions, file contents, debugging traces, user preferences. Traditional memory search returns the full content of every result, burning through tokens quickly and leaving little room for the actual task.

Infinite Context solves this with a two-step retrieval pattern called Progressive Disclosure. Instead of loading everything at once, you first retrieve a lightweight index of matching memories, then selectively load only the ones you need. The result: agents can preview dozens or even hundreds of memories in the token cost of just a few full results.

How It Works

Progressive disclosure splits memory retrieval into two explicit steps:

Token Savings

The progressive disclosure pattern dramatically reduces token usage compared to loading full content for every search result.

When to Use

Choose the right retrieval approach based on your scenario:

Quick Example

import { RekallAgent } from '@rekall/agent-sdk';

const agent = new RekallAgent({ apiKey: 'rk_...', agentId: 'agent_123' });

// Step 1: Get lightweight index
const index = await agent.recallIndex('project requirements');
console.log(`Found ${index.index.length} memories using ${index.indexTokens} tokens`);

// Step 2: Review snippets, select relevant
const relevant = index.index
  .filter(entry => entry.score > 0.7)
  .map(entry => entry.id);

// Step 3: Load full content for selected only
const memories = await agent.recallFull(relevant);

// Or use smartRecall for automatic handling
const auto = await agent.smartRecall('project requirements', {
  tokenBudget: 4000,
  progressive: true,
});

Observations

The memory_observe / observe() method captures tool outputs, file reads, API responses, and other external data your agent encounters during a session.

These observations are automatically indexed and become searchable via progressive disclosure in future sessions. This is the key mechanism for building up a searchable knowledge base over time — every piece of information your agent encounters can be stored cheaply, then retrieved efficiently later through the two-step index-then-load pattern.

Token Budget

Token budgets control how much of the context window memory retrieval is allowed to consume.

• •Default budget: 4,000 tokens
• •Configurable via preferences: Set rekall_token_budget in user preferences to adjust globally
• •Per-request control: The tokenBudget parameter on search controls how many index entries to return
• •Automatic management: smartRecall() automatically manages the budget across both the index and full-load steps

MCP Prompts

MCP clients (Claude Desktop, Cursor, Windsurf, etc.) can request built-in prompts that are injected at session start. These give the agent context on how to use memory efficiently from the very first message.

Next Steps