MEMOIR: A Training-Free Memory Lifecycle for Long-Horizon LLM Agents
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "MEMOIR: A Training-Free Memory Lifecycle for Long-Horizon LLM Agents".
Jane: Here's the gist of the paper. Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed The full retrieval stack improves evidence recall by 11.92 percentage points over bare cosine on LoCoMo MEMOIR proposes a training-free memory lifecycle that unifies judge-gated consolidation with retrospective attribution This approach addresses two coupled failure modes: prospective corruption where consolidation destroys evidence required by future queries, and retrospective staleness where retrieval favors stale memories The system operates on the principle of "attribute usefulness after use, and screen for preservation before compression" Pruning redundant reasoning tokens The proposed lifecycle involves a prospective screening gate to test whether answer-relevant evidence is preserved before compressing memory After each query, attribution estimates which retrieved memories supported the answer and updates their retrieval priority using TRACE Judge-gated merge changes the recall of text-based evidence from 58.82% to 59.08% and reduces the active store by 1.42% with zero observed losses in 100 audited merges The consolidation operator is deliberately merge-only, applying a condition where Apply(C, z) = 1 if c(C, z) ≥ τc and κ(C, z) ≥ τr How the training signal is built Retrospective TRACE estimates memory support via lexical overlap in the training-free default or leave-one-out counterfactual change in the TRACE-CF variant Importance then accumulates only after use, as wimp(mi) ← min(1, wimp(mi) + αfbτidi) The UCB exploration bonus couples these importance updates with count-based exploration to reduce feedback-driven retrieval lock-in via bt(mi) = λu/s log(1 + t)/(1 + ni) This mechanism ensures that multisession memory behaves as a Restless Multi-Armed Bandit (RMAB), handling staleness non-destructively Where the gains come from The full retrieval stack improves evidence recall by 11.92 percentage points over bare cosine on LoCoMo The confidence-gated autonomous feedback improves evidence recall@5 by 1.50 percentage points over the matched no-TRACE retriever The full stack remains 11.92 points above bare cosine, while confidence-gated feedback adds 1.50 points overall without reference answers or reranker training Certified merge changes text-grounded evidence recall from 58.82% to 59.08%, reducing the active store by 1.42% with zero observed losses in 100 independently audited merges Lifecycle consistency and safety The lifecycle mechanisms share one invariant: an item may retain retrieval credit only while it preserves the evidence that earned that credit SUPERSEDE retires stale support reversibly and transfers discounted credit because the valid fact has changed, whereas UCB continues to probe low-exposure items to prevent lock-in The primary 100-operation certified merge reduces active-store tokens by 1.42% while raising fixed-budget Rec@5 from 58.82% to 59.08%, with zero audited losses The overall results outline a practical deployment path where the lifecycle reduces to one principle: attribute usefulness after use, and screen for preservation before compression The one-sided 95% upper loss-rate bound of 2.95% characterizes a conservative retrieval-preservation screen rather than an end-toend answerability guarantee MEMOIR frames long-horizon agent memory as a coupled lifecycle: screen prospectively and attribute retrospectively On 1,905 LoCoMo QA pairs, the full retrieval stack gains 11.92 points over bare cosine In matched Llama-3.1-70B experiments, MEMOIR reaches 58.70%/61.20% accuracy/recall on LoCoMo and 81.20%/85.10% on LongMemEval-S at 0.2k tokens and 1.18 s latency Prospectively, 100 empirical judge-certified merges reduce activestore tokens by 1.42% while raising fixed-budget Rec@5 from 58.82% to 59.08%, with 0/100 audited losses Overall, MEMOIR reduces the lifecycle to one closed loop: attribute usefulness after use, and certify preservation before compression The results are concentrated on ten LoCoMo stores The final accounting uses P = 1,120 pre-consolidation answerable probes and reports exact lost-probe numerators in Table 17 The paper is 2611.07631.
Tom: Page 5 --- MEMOIR (Memory Evidence via Merge-certified, Online Importance Revision) Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed Existing systems treat these as separate problems: memories are written prospectively with static importance scores that are never revised after use This creates two coupled failure modes: consolidation can silently destroy evidence that future queries require, while retrieval continues to favor stale memories Page 9 --- Table 3 separates credit, exploration, and feedback quality using store-level inference TRACE without UCB gains only 0.42 points (CI
−0.12, 0.96: , p =.1420; 6/10 positive folds) The raw-self UCB condition gains 1.13 points (CI
0.21, 2.05: , p =.0210; 9/10 folds) Verified feedback gains 1.50 points (CI
0.41, 2.59: , p =.0082) and oracle multi-hop gains are 8.05 points Page 6 --- Figure 1: MEMOIR architecture The online path retrieves memories with TRACE-UCB and applies a closed-form retrospective update from observed outcomes During periodic offline sleep, MEMOIR proposes a core MERGE or optional SUPERSEDE operation, commits it only when all screening gates pass Page 8 --- Algorithm 1 summarizes the full closed-loop procedure referenced in the main paper’s Methodology section Online path (lines 1–6, executed per query) Offline path (lines 7–19, executed only during sleep) Page 4 --- MEMOIR optimizes two coupled objectives: improve evidence recall through retrospective credit while allowing only judge-gated lifecycle changes intended to preserve answerability or explicitly retire stale evidence with provenance Credit-Preservation Lifecycle Invariant Page 5 --- The bridge is that attribution provides online estimates of support sets for consolidation, while accepted merges preserve and transfer the credit ledger In short, the system consolidates what the screening procedure accepts as redundant and up-weights what is estimated to be useful Page 5 --- The goal is not simply better retrieval, but a memory lifecycle in which compression and credit assignment are mutually consistent Page 9 --- TRACE-UCB couples these with count-based exploration to mitigate exposure lock-in The final score is St(q, mi) = SRQ(q, mi) + bt(mi) Page 5 --- The online write path requires no LLM call and takes 6.8ms median latency This path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one Page 9 --- The online path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one Page 5 --- The online write path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one
Title and authors: Tom: So, we're talking about this paper today titled "MEMOIR: A Training-Free Memory Lifecycle for Long-Horizon LLM Agents". It sounds like they're tackling that big problem of long-term AI agents needing to remember useful stuff across many sessions.
Jane: That makes sense. Basically, existing systems write memories prospectively with fixed importance scores that never change once a memory is written, which creates two bad things: consolidation can accidentally delete evidence needed later, and retrieval just keeps picking old stuff even if it's no longer relevant.
Lu: Exactly. This paper proposes MEMOIR as a way to unify the process of consolidating memory with figuring out how useful that memory actually is retrospectively after use, using something called judge-gated consolidation and retrospective attribution.
Meng: I'm curious about what "training-free" means here. Does it mean they aren't training anything new during the retrieval process? Because right now, I see a lot of systems relying on fine-tuning for better recall.
Lalam: It says they achieve this without needing any new training during the online write or retrieval path, and they use something called TRACE to estimate memory support after a query happens.
The paper's summary: Tom: So, the core idea is that before you compress memory for storage, there's a screening gate that tests if the evidence relevant to an answer is actually preserved, and after a query, attribution estimates which retrieved memories helped and updates their priority.
Jane: That sounds like they are trying to fix both failures at once: prospective corruption where consolidation destroys future needs, and retrospective staleness where retrieval favors old information.
Lu: They frame it as being able to screen for preservation before compression, and then updating importance based on how useful the memory was after use. This is what they call "attribute usefulness after use, and screen for preservation before compression".
Meng: But if we're talking about practical deployment at scale, how do you handle the fact that this process runs every time an agent queries something? Does it slow things down significantly?
Lalam: The online write path actually takes six point eight milliseconds on average, and it uses lexical TRACE for attribution estimation without needing any additional LLM calls during the retrieval path.
The paper's improvements: Tom: One of the big numbers I see is that this full retrieval stack improves evidence recall by eleven point nine two percentage points over bare cosine similarity on LoCoMo tests. That’s a solid lift from just using simple cosine similarity.
Jane: And they also show that the confidence-gated autonomous feedback adds another one point five over the matched no-TRACE retriever for recall at five percent accuracy, which is pretty significant when you're talking about finding facts.
Lu: The paper details two main ways they build their training signal: TRACE-FB uses lexical overlap, and TRACE-CF uses counterfactual attribution, which measures support by leaving out one piece of the answer to see if it changes.
Meng: I’m looking at how they handle lock-in. They mention a UCB style exploration bonus to couple these importance updates with count based exploration, which is designed to prevent the system from getting stuck on frequently retrieved memories.
Lalam: They also have this consolidation operator that is deliberately merge only, applying a condition where it applies if the consolidation score is greater than tau c and the retrieval score is greater than tau r.
Conclusion: Tom: So to wrap up, the main implication here is that we can have a memory lifecycle that isn't static, but instead depends on that judge-gated screening process to ensure answerability before compressing, while using retrospective attribution to update priorities based on actual use.
Jane: The paper frames this whole thing as a closed loop where attribute usefulness after use and screen for preservation before compression are the two principles guiding the memory lifecycle.
Lu: It seems like the focus is less on just better retrieval and more on making sure that compression and credit assignment are actually consistent, which is a big conceptual step for how we think about long-horizon agent memory.
Meng: From an engineering standpoint, those certified merge changes reduce active store tokens by one point four two percent in one hundred audited merges without any observed losses at all. That's concrete performance data that shows the system is actually working as intended on a set of tests.
Lalam: And they show that this conservative screen, which has a ninety five percent upper loss rate bound of two point nine five percent, is really about being a conservative retrieval-preservation screen rather than guaranteeing answerability end to end.
Mayank Anand, Subham Raj, Aman Vaibhav Jha, Sriparna Saha
Indian Institute of Information Technology Allahabad · Indian Institute of Technology Patna
cs.IR
Submitted: 2026-11-05
Updated: 2026-11-05
Code: https://github.com/AnandMayank/MEMOIR
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 89/100
The gist: Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed The full retrieval stack improves evidence recall by 11.92 percentage…
Key concepts
- Attribute Usefulness After Use
- This core principle dictates the memory lifecycle: an item's value is assessed after it has been used for a query. Compression or consolidation only occurs if the evidence is deemed useful based on this post-use evaluation, preventing the destruction of necessary information.
- TRACE Judge-gated Merge
- This mechanism uses attribution estimates from past queries to determine which retrieved memories supported an answer. It acts as a prospective screening gate, ensuring that only evidence deemed relevant is preserved during consolidation, thus improving recall without needing retraining.
- UCB Exploration Bonus
- To prevent retrieval lock-in on frequently used memories, this bonus couples importance updates with count-based exploration. It actively encourages the system to retrieve under-exposed or less-seen memories, ensuring a diverse and robust memory store.
Terminology
Summary
Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed The full retrieval stack improves evidence recall by 11. p. 0 MEMOIR: A Training-Free Memory Lifecycle for Long-Horizon LLM Agents
Mayank Anand1
, Subham Raj2∗
, Aman Vaibhav Jha2
, Sriparna Saha2
1
Indian Institute of Information Technology Allahabad
2
Indian Institute of Technology Patna
Corresponding author: subham_2221cs25@iitp. in
Abstract
Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when
needed.Where the paper says it
attribute usefulness after use, and screen for preservation before compression
Pruning redundant reasoning tokens
The proposed lifecycle involves a prospective screening gate to test whether answer-relevant evidence is preserved before compressing memory After each query, attribution estimates which retrieved memories supported the answer and updates their retrieval priority using TRACE Judge-gated merge changes the recall of text-based evidence from 58.82% to 59.08% and reduces the active store by 1. p. 0 Judge-gated
merge changes the recall of text-based evidence from 58. 08%, reduces the active store by 1.Where the paper says it
How the training signal is built
Retrospective TRACE estimates memory support via lexical overlap in the training-free default or leave-one-out counterfactual change in the TRACE-CF variant Importance then accumulates only after use, as wimp(mi) ← min(1, wimp(mi) + αfbτidi) The UCB exploration bonus couples these importance updates with count-based exploration to reduce feedback-driven retrieval lock-in via bt(mi) = λu/s log(1 + t)/(1 + ni) This mechanism ensures that multisession memory behaves as a Restless Multi-Armed Bandit (RMAB), handling staleness non-destructively
Where the gains come from
The full retrieval stack improves evidence recall by 11.92 percentage points over bare cosine on LoCoMo The confidence-gated autonomous feedback improves evidence recall@5 by 1. p. 0 On 1,905 deduplicated LoCoMo QA pairs, the
full retrieval stack improves evidence recall@5 by 11. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. Confidence-gated autonomous feedback improves evidence recall@5 by 1. 50
percentage points over the matched no-TRACE retriever,
while reference-answer feedback reaches a 1. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. 50 points overall without reference answers
or reranker training; a deployment-unavailable oracle adds
1. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. 50 points overall without reference answers
or reranker training; a deployment-unavailable oracle adds
1. Judge-gated
merge changes the recall of text-based evidence from 58. 08%, reduces the active store by 1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Lifecycle consistency and safety
The lifecycle mechanisms share one invariant: an item may retain retrieval credit only while it preserves the evidence that earned that credit SUPERSEDE retires stale support reversibly and transfers discounted credit because the valid fact has changed, whereas UCB continues to probe low-exposure items to prevent lock-in The primary 100-operation certified merge reduces active-store tokens by 1.42% while raising fixed-budget Rec@5 from 58. p. 7 Prospectively, 100 empirical judge-certified merges reduce activestore tokens by 1. 42% while raising fixed-budget Rec@5
from 58. Prospectively, 100 empirical judge-certified merges reduce activestore tokens by 1. 42% while raising fixed-budget Rec@5
from 58. Together,
these results outline a practical deployment path: no additional LLM call on the online write path, with merge screening confined to periodic offline consolidation. The lifecycle
reduces to one principle: attribute usefulness after use, and
screen for preservation before compression.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
MEMOIR frames long-horizon agent memory as a coupled lifecycle: screen prospectively and attribute retrospectively On 1,905 LoCoMo QA pairs, the full retrieval stack gains 11.92 points over bare cosine In matched Llama-3. p. 5 The full stack remains 11. 92 points above bare
cosine, while Table 3 reports the matched feedback/UCB
effects. In matched Llama-3. 1-70B experiments, MEMOIR
reaches 58. 1-70B experiments, MEMOIR
reaches 58. 20% accuracy/recall on LoCoMo and
81. 3 Long-Horizon Memory Benchmarks
We evaluate MEMOIR primarily on LoCoMo, a longconversation benchmark spanning up to 32 sessions per dialogue with single-hop, multi-hop, temporal, and adversarial
query categories (Maharana et al. We also consider
LongMemEval-S, which targets memory abilities such as Primary LoCoMo inference uses 10,000 bootstrap
resamples over the ten stores; question-level bootstrap is a
sensitivity analysis, and leave-one-store-out folds test consistency. Consolidation is evaluated using fixed-budget, textgrounded recall, active-store tokens, and an independent
model-family audit. 8 ms LLM-free write path. Prospectively, 100 empirical judge-certified merges reduce activestore tokens by 1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 5 ---
MEMOIR (Memory Evidence via Merge-certified, Online Importance Revision) Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed Existing systems treat these as separate problems: memories are written prospectively with static importance scores that are never revised after use This creates two coupled failure modes: consolidation can silently destroy evidence that future queries require, while retrieval continues to favor stale memories
--- Page 9 ---
Table 3 separates credit, exploration, and feedback quality using store-level inference TRACE without UCB gains only 0.42 points (CI [−0. p. 0 On 1,905 deduplicated LoCoMo QA pairs, the
full retrieval stack improves evidence recall@5 by 11. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. 1420; 6/10 positive folds). The raw-self UCB condition gains 1. On 1,905 deduplicated LoCoMo QA pairs, the
full retrieval stack improves evidence recall@5 by 11. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. 84-point oracle gain; both are positive
in all folds. Verified and oracle multi-hop gains are 7. On 1,905 deduplicated LoCoMo QA pairs, the
full retrieval stack improves evidence recall@5 by 11. 92 percentage points over bare cosine, with confidence-gated feedback adding 1. 84-point oracle gain; both are positive
in all folds. Verified and oracle multi-hop gains are 7.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 6 ---
Figure 1: MEMOIR architecture The online path retrieves memories with TRACE-UCB and applies a closed-form retrospective update from observed outcomes During periodic offline sleep, MEMOIR proposes a core MERGE or optional SUPERSEDE operation, commits it only when all screening gates pass
--- Page 8 ---
Algorithm 1 summarizes the full closed-loop procedure referenced in the main paper’s Methodology section Online path (lines 1–6, executed per query) Offline path (lines 7–19, executed only during sleep)
--- Page 4 ---
MEMOIR optimizes two coupled objectives: improve evidence recall through retrospective credit while allowing only judge-gated lifecycle changes intended to preserve answerability or explicitly retire stale evidence with provenance Credit-Preservation Lifecycle Invariant
--- Page 5 ---
The bridge is that attribution provides online estimates of support sets for consolidation, while accepted merges preserve and transfer the credit ledger In short, the system consolidates what the screening procedure accepts as redundant and up-weights what is estimated to be useful
--- Page 5 ---
The goal is not simply better retrieval, but a memory lifecycle in which compression and credit assignment are mutually consistent
--- Page 9 ---
TRACE-UCB couples these with count-based exploration to mitigate exposure lock-in The final score is St(q, mi) = SRQ(q, mi) + bt(mi)
--- Page 5 ---
The online write path requires no LLM call and takes 6.8ms median latency This path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one
--- Page 9 ---
The online path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one
--- Page 5 ---
The online write path requires no additional LLM call on the online write or retrieval path, with lexical TRACE being the only one
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Memory lifecycle unification allows for
attribute usefulness after use, and screen for preservation before compression,
which means memory consolidation is not static but depends on a judge-gated screening process that tests whetheranswer-relevant evidence is preserved
before merging. -
The system can perform online writing with no additional LLM call and low latency while simultaneously updating memory importance through
TRACE (Trainingfree Retrieval Attribution and Credit Estimation)
using either lexical overlap or counterfactual attribution, enabling atraining-free architecture.
-
The agent can mitigate retrieval lock-in by using the UCB exploration bonus, which
encourages retrieval of under-exposed memories
to reduce the risk that frequently retrieved memories crowd out useful but lessseen evidence. -
Memory consolidation can be made conservative by employing judge-gated merge screening where a candidate cluster is replaced only when three gates pass:
c(C, z) ≥ τc, κ(C, z) ≥ τr, z ≤ η,
ensuring that the merge islossless
by preserving all unique information. -
The system can handle outdated information reversibly through the SUPERSEDE operator which allows for
temporal updates and stale evidence mitigation
by retiring older items and transferring discounted credit:wimp(mn) ← max(wimp(mn), λswimp(mo)), λs = 0.5.
-
The agent can perform session-level fact extraction using the DIGEST operator, which is an
additive, session-level synthesis operator
that generates facts independently answerable rather than burying them in a paragraph.
Abstract
Long-horizon LLM agents require memory systems that preserve useful information across sessions and retrieve it when needed. Existing systems treat these as separate problems: memories are written prospectively with static importance scores that are never revised after use. This creates two coupled failure modes: consolidation can silently destroy evidence that future queries require, while retrieval continues to favor stale memories. We propose MEMOIR (Memory Evidence via Merge-certified, Online Importance Revision), a training-free memory lifecycle that unifies judge-gated consolidation with retrospective attribution. Before compressing memory, screening gates test whether answer-relevant evidence is preserved. After each query, attribution estimates which retrieved memories supported the answer and updates their retrieval priority. A UCB-style exploration bonus additionally encourages retrieval of under-exposed memories, reducing the risk that frequently retrieved memories crowd out useful but less-seen evidence. On 1,905 deduplicated LoCoMo QA pairs, the full retrieval stack improves evidence recall@5 by 11.92 percentage points over bare cosine, with confidence-gated feedback adding 1.50 points overall without reference answers or reranker training; a deployment-unavailable oracle adds 1.84 points overall and 8.05 points on multi-hop. Judge-gated merge changes the recall of text-based evidence from 58.82% to 59.08%, reduces the active store by 1.42%, and incurs 0 observed losses in 100 independently audited merges, with a one-sided 95% upper loss-rate bound of 2.95%. Together, these results outline a practical deployment path: no additional LLM call on the online write path, with merge screening confined to periodic offline consolidation. The lifecycle reduces to one principle: attribute usefulness after use, and screen for preservation before compression.
Sources
- Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents
- Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
- MemOS: An Operating System for Memory-Augmented Generation (MAG) in Large Language Models
- Zep: A Temporal Knowledge Graph Architecture for Agent Memory
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
- HiMPO: Hindsight-Informed Memory Policy Optimization for Less-Entangled Credit in Long-Horizon Agents
- Are We Ready For An Agent-Native Memory System?
Related papers
- UniRank: Benchmarking Ranking Models for Unified Sequential Modeling and Feature Interaction
- Revisiting Text Ranking in Deep Research
- GateSID: One Gate to Balance Semantic and Collaborative Signals across the Item Lifecycle
- SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce
- GINGER: Spicing Up Graph-Based Vector Search
- Subtraction Gets You More: Gap-Aware Retrieval for Multimodal Multi-Hop QA