Decoded
← Feed
arXiv:2607.21503AIIR23 jul 2026

Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

1 min readHN ×7PDF ↗
SINGLE-SESSION-USER100%KNOWLEDGE-UPDATE100%TEMPORAL-REASONING100%SINGLE-SESSION-ASSISTANT87.5%0100% LONGMEMEVAL CATEGORY ACCURACY
fig. longmemeval category accuracyPerfect on single-session recall, weakest across sessions

Agent memory should be a managed lifecycle, not just a database, to cut cost and forgetting.

92.0%

LongMemEval overall accuracy

Uses the smaller gpt-5-mini as both answer model and judge.

How it works

  1. Agents drown in history

    Re-sending the full chat every turn makes token cost grow quadratically. Crude summaries drop key facts and hurt accuracy.

  2. Five-part context lifecycle

    Design memory per agent, extract structure, scope retrieval by user and org, prefetch ahead, and compact with validation.

  3. High memory benchmark scores

    Synap scores 92.0% on LongMemEval and 93.2% on LoCoMo. Multi-session reasoning stays weakest at 75.2%.

configuresstored memoryfit budgetArchitectingIngestingScopingAnticipatingValidatedcompactingconfiguresstored memoryfit budgetArchitectingIngestingScopingAnticipatingValidatedcompacting

What they found

Which comes out on top on LongMemEval category accuracy?

Bounded context savings over time

try it

Drag conversation length up and watch token savings versus full-append grow.

Conversation lengthToken savingsShort chats: small savingsLong chats: ~96% savedroughly 90% savings
Token savings
medium
here
Short chats: small savings
Full-append cost grows quadratically, so a fixed budget saves more the longer agents talk.Illustrative model, not the paper's data

Terms

6 terms

Full-append Re-sending the entire conversation history to the model on every turn.

Like a good executive assistant: files notes carefully, pulls the right folder before a meeting, and hands you a one-page brief checked against the originals.

Where it breaks
An assistant understands meaning; the system relies on proprietary, unpublished checks.

Limits

  • Core mechanisms are proprietary, so results cannot be independently reproduced in detail.
  • No latency or token-cost measurements; savings figures are illustrative math.
  • Competitor scores are self-reported with different setups, not a controlled comparison.
From the paper