@inproceedings{banerjee-etal-2026-apex-em,
title = "{APEX}-{EM}: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience Replay",
author = "Banerjee, Pratyay and
Moshtaghi, Masud and
Chadha, Ankit",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1996/",
pages = "42547--42568",
ISBN = "979-8-89176-462-0",
abstract = "LLM agents rerun full reasoning for every task, even one they solved moments earlier. We introduce APEX-EM, a non-parametric experience memory that stores complete procedural-episodic traces in a typed Procedural Knowledge Graph (PKG) and retrieves them through three channels: semantic search, structural-signature matching over abstract operation sequences, and graph traversal. A Plan-Retrieve-Generate-Iterate-Ingest (PRGII) workflow produces, quality-gates, and commits experiences, indexing both successes and failures so the agent learns what to reuse and what to avoid. No weights change during deployment. We evaluate on five benchmarks: BigCodeBench, KGQAGen-10k, HLE, Lifelong Agent Bench, and ALFWorld. Because prior work uses different backbones, we base our claims on same-backbone comparisons that hold model capability fixed. On held-out BigCodeBench transfer with a shared GPT-4o backbone, APEX-EM gains +7.6pp over the no-memory baseline, which is 3.3x MemRL{'}s +2.3pp under the identical setup. On Lifelong Agent Bench with a shared GPT-4o-mini backbone, APEX-EM gains +1.4pp (OS) and +1.0pp (DB) cumulative success. On KGQAGen-10k, frozen memory transfers to a blind 1{,}079-question test split at 73.7{\%} versus 42.0{\%} with no memory, approaching an oracle that is handed the ground-truth subgraph (84.9{\%}). Across three Opus scales the memory gain stays at +27pp to +32pp, so it adds to model capability rather than substituting for it. Component analysis shows no single mechanism dominates: teacher feedback is negligible for code but adds +10.3pp on structured queries, structural signatures give 3.3 times the transfer of semantic-only retrieval, and within-epoch iteration recovers most of the gain when rich feedback is unavailable. These results argue for modular memory composed per domain."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="banerjee-etal-2026-apex-em">
<titleInfo>
<title>APEX-EM: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience Replay</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pratyay</namePart>
<namePart type="family">Banerjee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Masud</namePart>
<namePart type="family">Moshtaghi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ankit</namePart>
<namePart type="family">Chadha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>LLM agents rerun full reasoning for every task, even one they solved moments earlier. We introduce APEX-EM, a non-parametric experience memory that stores complete procedural-episodic traces in a typed Procedural Knowledge Graph (PKG) and retrieves them through three channels: semantic search, structural-signature matching over abstract operation sequences, and graph traversal. A Plan-Retrieve-Generate-Iterate-Ingest (PRGII) workflow produces, quality-gates, and commits experiences, indexing both successes and failures so the agent learns what to reuse and what to avoid. No weights change during deployment. We evaluate on five benchmarks: BigCodeBench, KGQAGen-10k, HLE, Lifelong Agent Bench, and ALFWorld. Because prior work uses different backbones, we base our claims on same-backbone comparisons that hold model capability fixed. On held-out BigCodeBench transfer with a shared GPT-4o backbone, APEX-EM gains +7.6pp over the no-memory baseline, which is 3.3x MemRL’s +2.3pp under the identical setup. On Lifelong Agent Bench with a shared GPT-4o-mini backbone, APEX-EM gains +1.4pp (OS) and +1.0pp (DB) cumulative success. On KGQAGen-10k, frozen memory transfers to a blind 1,079-question test split at 73.7% versus 42.0% with no memory, approaching an oracle that is handed the ground-truth subgraph (84.9%). Across three Opus scales the memory gain stays at +27pp to +32pp, so it adds to model capability rather than substituting for it. Component analysis shows no single mechanism dominates: teacher feedback is negligible for code but adds +10.3pp on structured queries, structural signatures give 3.3 times the transfer of semantic-only retrieval, and within-epoch iteration recovers most of the gain when rich feedback is unavailable. These results argue for modular memory composed per domain.</abstract>
<identifier type="citekey">banerjee-etal-2026-apex-em</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1996/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42547</start>
<end>42568</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T APEX-EM: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience Replay
%A Banerjee, Pratyay
%A Moshtaghi, Masud
%A Chadha, Ankit
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F banerjee-etal-2026-apex-em
%X LLM agents rerun full reasoning for every task, even one they solved moments earlier. We introduce APEX-EM, a non-parametric experience memory that stores complete procedural-episodic traces in a typed Procedural Knowledge Graph (PKG) and retrieves them through three channels: semantic search, structural-signature matching over abstract operation sequences, and graph traversal. A Plan-Retrieve-Generate-Iterate-Ingest (PRGII) workflow produces, quality-gates, and commits experiences, indexing both successes and failures so the agent learns what to reuse and what to avoid. No weights change during deployment. We evaluate on five benchmarks: BigCodeBench, KGQAGen-10k, HLE, Lifelong Agent Bench, and ALFWorld. Because prior work uses different backbones, we base our claims on same-backbone comparisons that hold model capability fixed. On held-out BigCodeBench transfer with a shared GPT-4o backbone, APEX-EM gains +7.6pp over the no-memory baseline, which is 3.3x MemRL’s +2.3pp under the identical setup. On Lifelong Agent Bench with a shared GPT-4o-mini backbone, APEX-EM gains +1.4pp (OS) and +1.0pp (DB) cumulative success. On KGQAGen-10k, frozen memory transfers to a blind 1,079-question test split at 73.7% versus 42.0% with no memory, approaching an oracle that is handed the ground-truth subgraph (84.9%). Across three Opus scales the memory gain stays at +27pp to +32pp, so it adds to model capability rather than substituting for it. Component analysis shows no single mechanism dominates: teacher feedback is negligible for code but adds +10.3pp on structured queries, structural signatures give 3.3 times the transfer of semantic-only retrieval, and within-epoch iteration recovers most of the gain when rich feedback is unavailable. These results argue for modular memory composed per domain.
%U https://aclanthology.org/2026.findings-emnlp.1996/
%P 42547-42568
Markdown (Informal)
[APEX-EM: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience Replay](https://aclanthology.org/2026.findings-emnlp.1996/) (Banerjee et al., Findings 2026)
ACL