---

title: "Summarization Drift" type: glossary tags: [#memory, #reasoning] created: 2025-07-15 updated: 2025-07-15 status: complete

Summarization Drift

Summarization drift is a memory failure mode in which iterative compression of an agent's interaction history progressively discards detail, producing a summary that diverges from the original events and introduces factual distortions into the agent's belief state.

Overview

Summarization drift occurs when an agent — or its memory management layer — repeatedly compresses a growing interaction history to keep it within a manageable size. Each compression pass is lossy: details are dropped, nuances are collapsed, and the compression artifact becomes the new input for the next compression. Over successive cycles, the cumulative distortion can be severe enough that the agent's working memory no longer accurately reflects what actually happened.

This is one of the primary failure modes within the context-resident compression family of memory mechanisms, and it is directly relevant to widely-used tools like Claude Code and Kiro CLI, both of which perform automatic context compression when sessions grow long. Practitioners commonly observe that these tools behave increasingly erratically in extended sessions — not because of model degradation, but because the compressed memory no longer faithfully represents the session history.

Summarization drift is distinct from staleness (where external facts change and memory does not update) and from over-generalization (where a narrow lesson is incorrectly applied broadly). It is specifically the artifact of the compression mechanism itself introducing inaccuracy.

How It Works

Raw History (T=1):  [Event A] [Event B] [Event C] [Event D]
                          │
                   Compression Pass 1
                          │
Summary (T=2):     [AB-summary] [CD-summary]   ← some detail lost
                          │
                   Compression Pass 2
                          │
Summary (T=3):     [ABCD-summary]              ← more detail lost
                          │
                   Compression Pass 3
                          │
Summary (T=4):     [Distorted-summary]         ← original events unrecoverable

Each pass is a one-way operation: once detail is dropped, it cannot be recovered from the compressed artifact alone. The agent then reasons on the distorted summary as if it were ground truth.

Mitigation: Keep raw episodic records alongside summaries. The raw log serves as an archival ground truth that agents can retrieve when a summary proves insufficient or contradictory. Versioning summaries with timestamps also allows agents to identify when a summary was generated and to flag it as potentially stale.

Key Properties / Characteristics

• Caused by iterative lossy compression of interaction history
• Failure is cumulative and accelerating — each pass compounds prior distortions
• The agent has no internal signal that drift has occurred; it treats distorted summaries as ground truth
• Observable symptom: agent responses become more generic, less grounded, and more likely to contradict known facts in long sessions
• Mitigated by preserving raw episodic records and versioning compressed artifacts
• Related to but distinct from attention dilution, which is a retrieval failure rather than a storage distortion

Strengths & Limitations

Strengths

• N/A — summarization drift is a failure mode, not a design choice

Limitations

• Difficult to detect without comparing summary content to raw records
• Particularly insidious in tools that perform automatic compression invisibly to the user
• No widely standardized mitigation pattern exists across agent frameworks

Notable Uses / Applications

• Observed in Claude Code and Kiro CLI during long coding sessions; standard practitioner workaround is to start a new thread rather than extend the existing session
• Discussed in the context of the Write-Manage-Read Loop: drift results from a Manage phase that compresses without preserving raw records
• Motivates the builder guidance to keep raw episodic records as a distinct archive alongside any derived summaries

Source Material

1. A Practical Guide to Memory for Autonomous LLM Agents — Primary source; describes summarization drift as a context-resident failure mode with practitioner examples.
2. Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers — Formal survey cataloguing summarization drift within the taxonomy of agent memory failure modes.

Related Pages

Is a subtype of: Memory Systems Occurs in: Write-Manage-Read Loop Contrasts with: Scratchpad See also: Grounding

Open Questions

• Can compression algorithms be designed that are provably lossless for task-relevant content while still achieving meaningful size reduction?
• How should agent frameworks signal to users when compression-induced drift is likely to have occurred?
• Is there a practical threshold of compression passes beyond which raw-record retrieval should be forced?

---

Page type: glossary | Status: complete