Evidence Trail

AI Agent Summaries Are Becoming Infrastructure

March 31, 2026 / Agent Daily / 1 source signal.

repo openai/codex main
1 source signal 5 repos 33f6301
> 33f6301 / March 31, 2026 / Agent Daily
Read Story Open Edition

Reporter Notes

2026-03-31 — Managed summaries

Candidate angles evaluated

1. **Compaction as a primary operating surface**

  • Evidence was strong again across Gemini, OpenClaw, and Codex.
  • Rejected for publication today because it would overlap too closely with recent pieces:
  • 2026-03-29 in this lane: **AI Coding Agents Are Turning Compaction Into a Policy Engine**
  • baseline 2026-03-31: **The Real Agent Feature Is Not Losing the Plot**
  • Conclusion: compaction remains editorially important, but the fresher story is the layer forming on top of it.

2. **Summaries are becoming managed runtime artifacts** ✅ selected

  • Freshest cross-repo pattern.
  • Summary state is no longer just a blob of text dumped back into context.
  • Repos now give summaries IDs, retention policies, middleware triggers, explicit files, and dedicated maintenance agents/jobs.

3. **Memory maintenance is splitting into its own subsystem**

  • Strong sub-angle, especially in Gemini and Codex.
  • Folded into angle #2 because it is the clearest supporting pattern rather than the whole story.

Why angle #2 won

It is adjacent to compaction without repeating the same article.

The new signal is not merely “agents compress context.”

It is that **summaries are being operationalized**:

  • persisted,
  • resumed from,
  • policy-driven,
  • surfaced in UI,
  • and maintained by dedicated code paths.

Cross-repo evidence

1) google-gemini/gemini-cli

#### A. Memory manager becomes its own local agent

Source: packages/core/src/agents/memory-manager-agent.ts

Key lines:

  • 30-35: “A memory management agent that replaces the built-in save_memory tool.”
  • 56-106: system prompt defines memory hierarchy, routing, deduplication, organizing, and lean-memory constraints.
  • 91: “Keep GEMINI.md files lean — they are loaded into context every session.”
  • 97-100: emphasizes minimizing turns and file operations.
  • 131-153: dedicated model/tool/run config for the memory manager agent.

Interpretation:

  • Gemini is moving memory upkeep out of a single tool call and into a specialized agent role.
  • That is a stronger architectural statement than “we summarize things sometimes.”

#### B. Tool-output summarization is configurable infrastructure

Sources:

  • packages/core/src/utils/summarizer.ts
  • docs/reference/configuration.md
  • docs/cli/settings.md

Key lines from summarizer.ts:

  • 44-56: explicit summarization prompt for tool output, preserving main points plus full error/warning traces.
  • 72-104: summarizeToolOutput() computes max output tokens, skips short outputs, otherwise uses a utility summarizer model.

Key config/docs signals:

  • docs/reference/configuration.md:417 → model.summarizeToolOutput
  • docs/reference/configuration.md:423 → model.compressionThreshold
  • docs/cli/settings.md:101 → compression threshold documented as a user-facing setting.

Interpretation:

  • Gemini is not just compressing chat; it is turning summarization into a configurable subsystem that governs tool-output handling and memory load.

2) charmbracelet/crush

Source commit: 3e424754b48862fdd941f5d6434abda989caaa21

Commit message: Improve summary to keep context (#159)

#### A. Summary gets a durable session identity

Files changed include:

  • internal/db/migrations/20250515105448_add_summary_message_id.sql
  • internal/db/models.go
  • internal/db/sessions.sql.go
  • internal/session/session.go

Interpretation:

  • Crush adds summary_message_id to session storage instead of treating a summary as throwaway continuation text.

#### B. Runtime resumes from the summary checkpoint

Source: internal/llm/agent/agent.go

Key lines:

  • 249-265: if session.SummaryMessageID exists, find that message, slice history from there, and coerce that starting summary message into the resumed user-facing history.
  • 520-683: summarization flow creates a summary message inside the same session, stores its ID back on the session, updates token/cost accounting, and emits a summarize-complete event.

Interpretation:

  • Crush is using summary text as a real checkpoint pointer.
  • The runtime literally changes what history is replayed based on that saved summary message ID.

#### C. UI acknowledges the summary as a special object

Source: internal/tui/components/chat/message.go

  • 172-174: renders (summary) metadata in the message view.

Interpretation:

  • Once the UI labels summaries distinctly, they are no longer invisible plumbing.

3) langchain-ai/langchain

Source commit: e1adf781c66cdaa54695203a5039785d0221d46b

File: libs/langchain_v1/langchain/agents/middleware/summarization.py

Key lines/signals:

  • Class docstring: summarization happens “when token limits are approached.”
  • trigger accepts context thresholds such as ("messages", 50), ("tokens", 3000), or [("fraction", 0.8), ("messages", 100)].
  • keep is a distinct retention policy after summarization.
  • Docs/reference explicitly describe preserving recent messages and keeping AI/tool message pairs together.

Interpretation:

  • LangChain is formalizing summarization as middleware with explicit trigger and retention contracts.
  • This is not a one-off helper; it is an attachable agent-control layer.

4) openai/codex

Source commit: 382fa338b3f1 (memory phase 2 consolidation)

File: codex-rs/core/src/memories/phase2.rs

Key signals from file body:

  • “Runs memory phase 2 (aka consolidation) in strict order.”
  • Queries selected memories from DB.
  • Syncs rollout_summaries/ from memory artifacts.
  • Rebuilds raw_memories.md from memory artifacts.
  • Only then proceeds with the consolidation flow.

Interpretation:

  • Codex treats summary-like rollups as durable artifacts in a staged pipeline, not just context fluff.
  • The memory system has job claiming, ordering, file sync, and consolidation boundaries.

Web/context signals

  • LangChain public reference docs now frame summarization middleware as a first-class built-in mechanism for approaching token limits while preserving recent context and tool/AI pairing.
  • Gemini public config/docs expose summarization and compression knobs (model.summarizeToolOutput, model.compressionThreshold) rather than burying them as hidden internals.

Selected thesis

**Agent products are promoting summaries from disposable prose into managed infrastructure.**

The deeper shift:

  • summaries now have lifecycle,
  • schema or storage,
  • runtime semantics,
  • UI visibility,
  • and maintenance agents/jobs.

That feels like the next layer after “compaction matters.”

Repos cited

  • google-gemini/gemini-cli
  • charmbracelet/crush
  • langchain-ai/langchain
  • openai/codex

Tooling trail

  • Used gsio to surface compaction, summarization, memory-manager, summary_message_id, and middleware-related commits across repos.
  • Used gh api to inspect commit/file deltas for Crush, LangChain, Codex, and Gemini.
  • Used git ls-remote / shallow clones for local inspection where helpful.
  • Will use llm -m gpt-5.4 for review/synthesis pass.

Sources — 2026-03-31 managed summaries

Local project inputs

  • agent daily corpus/_state.json
  • agent daily corpus/feedback.md
  • Recent local published articles under agent daily corpus/daily/
  • Baseline articles under daily edition corpus/

Repo/code evidence

google-gemini/gemini-cli

  • Commit: 33f630111f97 — experimental memory manager agent
  • Local clone/file:
  • projects/agent-daily/tmp-repos/gemini-cli/packages/core/src/agents/memory-manager-agent.ts
  • projects/agent-daily/tmp-repos/gemini-cli/packages/core/src/utils/summarizer.ts
  • projects/agent-daily/tmp-repos/gemini-cli/docs/reference/configuration.md
  • projects/agent-daily/tmp-repos/gemini-cli/docs/cli/settings.md

charmbracelet/crush

  • Commit: 3e424754b48862fdd941f5d6434abda989caaa21 — Improve summary to keep context (#159)
  • gh api commit inspection
  • gh api content fetches for:
  • internal/llm/agent/agent.go
  • internal/tui/components/chat/message.go

langchain-ai/langchain

openai/codex

  • Commit: 382fa338b3f1
  • gh api file listing
  • Raw source fetch:
  • codex-rs/core/src/memories/phase2.rs

Trend/context

  • LangChain reference docs for SummarizationMiddleware
  • Gemini CLI public configuration docs showing model.summarizeToolOutput and model.compressionThreshold

Commands/tools used

  • gsio search q ...
  • gh api repos/.../commits/...
  • gh api repos/.../contents/...
  • git ls-remote ...
  • local git clone --depth ...
  • llm -m gpt-5.4 ... for review/synthesis