---
id: "concept-anchored-iterative-summarization"
type: "concept"
source_timestamps: ["00:14:02", "00:14:25"]
tags: ["context-management", "llm-optimization"]
related: ["action-compress-context-iteratively", "claim-factory-compression-superiority"]
definition: "A context compression technique that merges newly truncated session history into a persistent, highly structured summary document to prevent information loss."
sources: ["s41-nvidia-open-sourced"]
sourceVaultSlug: "s41-nvidia-open-sourced"
originDay: 41
---
# Anchored Iterative Summarization

## Definition

A context-compression strategy that merges newly truncated session history into a persistent, highly structured summary document — preventing the loss of intent, decisions, and architectural state across long agent sessions.

## The Problem It Solves

Long-running agent sessions blow past LLM context windows. The two common alternatives both degrade fidelity:

1. **Naive full-summary** — ask the LLM to summarize the entire history. This produces a lossy "telephone game" effect, especially across multiple compression cycles. Critiqued in [[claim-factory-compression-superiority]] as the failure mode of [[entity-anthropic-d41]]'s Claude SDK approach.
2. **Opaque endpoint** — use a black-box compaction API (e.g., [[entity-openai-d41]]'s compact endpoint) where developers cannot verify what context was preserved.

## How It Works

Maintain a single **structured summary document** with explicit, immutable sections:

- **Session Intent** — original user/architect goal, never overwritten
- **File Modifications** — running ledger of files touched and why
- **Decisions Made** — architectural choices, with reasoning
- **Next Steps** — explicit forward-looking plan

When a context window threshold is hit:
1. The newly truncated span is summarized.
2. That summary is **explicitly merged** into the structured document — anchored into the appropriate section, not appended as free text.
3. The structured document persists forward; the raw history is dropped.

## Why It Works

The anchoring forces preservation of state that would otherwise drift. The agent always carries forward its original intent and the architectural decisions it has already locked in.

This is operationalized in [[action-compress-context-iteratively]] and benchmarked against native methods in [[claim-factory-compression-superiority]] (per [[entity-factory-ai-d41]]).

## Prerequisites

Requires understanding of [[prereq-context-window-mechanics]].

## See Also

- [[entity-factory-ai-d41]] — origin of the technique
- [[claim-factory-compression-superiority]] — the empirical claim
- [[action-compress-context-iteratively]] — the operational recipe