Compiled Memory: Not More Information, but More Precise Instructions for Language Agents

James Rhodes; George Kang

cs.AIMar 12, 2026

Compiled Memory: Not More Information, but More Precise Instructions for Language Agents

Authors: James Rhodes, George Kang

Organizations: AlphaBitCore, Inc.

Abstract

Existing memory systems for language agents address memory management: how to retrieve and page more information within a context budget. We address a complementary problem -- memory utility: what experience is worth keeping, and how it should change agent behavior. We present Atlas, a memory kernel that compiles accumulated task experience into an agent's instruction structure -- without fine-tuning, RAG, or human intervention. Memory is distillation, not storage; delivery is instruction rewriting, not context injection. Facts extracted from agent failures and successes are verified through a three-step promotion gate and delivered by rewriting the agent's system prompt with learned sub-bullets. On CUAD contract analysis, the evolved prompt improves GPT-4o token-level F1 by $+8.7$ pp and precision by $+12.5$ pp. On HotpotQA multi-hop QA, joint F1 improves $+3.16$ pp. An ablation isolates the mechanism's defining property -- the training signal constraint: the evolved prompt learns exactly what it is taught, and nothing more. Applied to Claude Sonnet~4.5 using the same evolved prompt -- compiled from GPT-4o errors, unchanged -- joint F1 improves $+2.31$ pp, with gains concentrating where Claude's stronger baseline leaves the most room -- confirming that the compiled knowledge is task-shaped, not model-shaped.

Figures & tables

All $\|\Delta\|>1$ pp: $p<10^{-7}$ . F1: $d=0.27$ .
Metric	Evo	Base	$\boldsymbol{\Delta}$
Token F1	$70.5\%$	$61.9\%$	$+8.7$ pp
Exact Match	$36.9\%$	$32.4\%$	$+4.5$ pp
Precision	$71.0\%$	$58.5\%$	$+12.5$ pp
Recall	$81.3\%$	$82.6\%$	$-1.3$ pp
FN Rate	$6.6\%$	$9.6\%$	$-3.0$ pp

Table 1: GPT-4o on CUAD (

n{=}3{,}000

paired observations).

Config	Docs	Facts	$\boldsymbol{\Delta}$ F1
50 docs, no dedup	50	4	$+6.7$ pp
200 docs, no dedup	200	13	$+5.5$ pp
400 + C1 only	400	20	$+1.0$ pp
C1+C2 (quality only)	1,140	7	$+4.4$ pp
C1–C5 (full)	1,140	23	$+9.1$ pp

Table 2: CUAD component ablation (

n{=}1

, 200 examples/variant).

Dim.	Example Rule
ANSWER	“Extract BOTH birth dates; do not infer from paragraph order”
ATTRIB.	“Include the sentence with the compared value, not intro sentences”
BOTH	“Extract from paragraph B AND cite sentences from both A and B”

Table 3: Multi-objective fact extraction categories (v3).

Metric	Evo	Base	$\boldsymbol{\Delta}$
Joint F1	$51.80\%$	$48.64\%$	$+3.16$ pp
Joint EM	$23.33\%$	$19.67\%$	$+3.67$ pp
Answer F1	$79.27\%$	$78.07\%$	$+1.21$ pp
Answer EM	$65.90\%$	$64.67\%$	$+1.23$ pp
Supp. Fact F1	$63.18\%$	$60.01\%$	$+3.17$ pp
Supp. Fact EM	$32.50\%$	$28.03\%$	$+4.47$ pp

Table 4: GPT-4o v3 on HotpotQA (

n{=}3{,}000

paired observations).

Metric	v2 $\Delta$	v3 $\Delta$	Shift
Answer F1	$+1.28$ pp	$+1.21$ pp	$-0.07$ (flat)
Answer EM	$+1.00$ pp	$+1.23$ pp	$+0.23$ (flat)
Supp. Fact F1	$+2.35$ pp	$+3.17$ pp	$+0.82$ pp
Supp. Fact EM	$+3.13$ pp	$+4.47$ pp	$+1.34$ pp
Joint F1	$+2.13$ pp	$+3.16$ pp	$+1.03$ pp

Table 5: v2 vs. v3 ablation: only the training signal changed.

Joint F1: $p{<}0.001$ , $d{=}0.067$ . Answer F1: $p{=}0.22$ (ns).
Metric	Cl. Evo	Cl. Base	Cl. $\Delta$	GPT $\Delta$
Joint F1	$54.93\%$	$52.61\%$	$+2.31$ pp	$+3.16$ pp
Joint EM	$28.90\%$	$26.43\%$	$+2.47$ pp	$+3.67$ pp
Answer F1	$82.01\%$	$81.69\%$	$+0.32$ pp	$+1.21$ pp
Supp. Fact F1	$65.47\%$	$63.10\%$	$+2.37$ pp	$+3.17$ pp
Supp. Fact EM	$40.10\%$	$36.73\%$	$+3.37$ pp	$+4.47$ pp

Table 6: Claude Sonnet 4.5 vs. GPT-4o v3 on HotpotQA.

n{=}6

runs

\times

500 examples

=3{,}000

paired observations (Claude);

n{=}6

(GPT-4o). Same evolved prompt for both—compiled exclusively from GPT-4o training errors.

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 373 papers

Muscle Memory for Agents: Compile not Merely Retrieve
Aug 10, 2026Pouya Ghiasnezhad Omran, Soujanya Lanka, Qin Zhang +1Conversational Memory LLM Agent Memory
Google Cloud FDE
Towards Autonomous Memory Agents
Feb 25, 2026Xinle Wu, Rui Zhang, Mustafa Anis Hussain +1Memory-Augmented Language Model Agents Memory Management for LLM Agents
National University of Singapore, Singapore
MemCompiler: Compile, Don't Inject -- State-Conditioned Memory for Embodied Agents
May 8, 2026Xin Ding, Xinrui Wang, Yifan Yang +9Embodied AI Memory-Augmented Agents
University of Science and Technology of China · Huazhong University of Science and Technology · Microsoft Research +2

Aug 10, 2026J/K move · Enter open · S save

Muscle Memory for Agents: Compile not Merely Retrieve

Pouya Ghiasnezhad Omran, Soujanya Lanka, Qin Zhang, Tanya Dixit

Google Cloud FDE

Conversational Memory LLM Agent Memory LLM Agents

Open paper AskPDF arXiv

Memory for LLM agents has converged on a single architectural pattern: store experience as text, embeddings, reflections, or rules; retrieve at inference time; let a general-purpose orchestrator interpret what to do. This paper argues that the pattern is the wrong default for personalization. We position Muscle Memory - the practice of compiling recurring user intent into purpose-built specialist agents - as a distinct memory paradigm from retrieval, and we argue that compilation is a better fit for the workloads where current assistants impose a multi-turn tax on their users: making them repeatedly correct format, depth, and scope to obtain a domain-appropriate answer. We support the position with a reference implementation and empirical evidence. The implementation is a four-phase pipeline (Harvest

\rightarrow

Analyze

\rightarrow

Augment

\rightarrow

Evaluate) that mines conversational history, separates behavioral from task patterns, and emits quality-gated executable compiled specialists with two-stage trigger matching. On 90 held-out scenarios across five user personas, the augmented assistant wins 32 of 36 cases where a specialist fires, an 88.9% win rate, with a +2.05 personalization gain and only a

-0.28

accuracy cost on a 1-4 scale. We discuss why compilation is better suited than retrieval in this regime, what the result implies for the broader memory design space, and what open problems remain.

Abstract

LLM Agent Memory

Explore similar work

Muscle Memory for Agents: Compile not Merely Retrieve

Towards Autonomous Memory Agents

MemCompiler: Compile, Don't Inject -- State-Conditioned Memory for Embodied Agents