---
title: 'InfoMem: Information-Centric Memory in LLMs'
url: https://www.emergentmind.com/topics/infomem
type: topic
---

# InfoMem: Information-Centric Memory in LLMs

InfoMem is an information-centric view of memory for large language models and agents in which the primary question is not merely how to store more context, but how to encode, preserve, retrieve, update, and, when necessary, delete information so that future tasks remain well supported under constraints of latency, bounded context, personalization, and maintainability. Recent work uses the term in closely related but not identical senses: as an information-theoretic memory bottleneck for long-horizon agents, as an answer-conditioned reward for training chunk-wise memory agents, and as a broader design target for modular memory subsystems that remain effective under noisy corpora, long documents, or multimodal retention [2602.07885] [2606.03329] [2605.15156].

## 1. Problem setting and scope

In the long-context setting studied by the reward mechanism called InfoMem, a document or corpus \(D\) is split into chunks \(D=\{c_1,\ldots,c_K\}\), and a chunk-wise agent updates a compact memory summary according to
\[
M_t=\pi_\theta(M_{t-1},c_t,x), \qquad \hat y=\pi_\theta(x,M_K),
\]
with each chunk about 5000 tokens and memory truncated to at most 1024 tokens [2606.03329]. The central difficulty is that the agent must preserve answer-relevant evidence while discarding noise, even though the evidence may be sparse and distributed.

A related formulation appears in MeMo, which defines a general knowledge integration mechanism
\[
(\Phi,f):\quad K=\Phi(\mathcal D),\qquad \hat a(q)=f(M_\theta,K,q),
\]
for a frozen Executive \(M_\theta\) with only black-box access, and instantiates \(K\) as the parameters of a separate memory model \(M_\varphi\) rather than a retrieval index or modified base-model weights [2605.15156]. In this setting, InfoMem refers to the problem of incorporating new, timely, domain-specific knowledge without modifying the Executive and without paying inference cost that grows with corpus size.

Other systems expand the scope from question answering over a fixed corpus to persistent, cross-session agent memory. MemFly treats the memory state \(M_t\) as a graph over Notes, Keywords, and Topics that evolves online under an information bottleneck objective [2602.07885]. Infini Memory treats persistent memory as topic-structured Markdown documents maintained over time through buffering, rewriting, routing, splitting, and merging [2606.10677]. AdaMem frames long-horizon personalization as a write-control problem in which an agent must learn what to remember for each role or user rather than remembering everything uniformly [2606.21144]. SwiftMem emphasizes low-latency retrieval over long conversational histories through query-aware temporal and semantic indexing [2601.08160]. MemLeak extends the setting further by showing that fact-level forgetting in multimodal agent memory is not equivalent to deleting a text record, because retained images and other residual representations can still support recovery of the deleted fact [2606.29788].

This suggests that InfoMem is best understood as a design space spanning bounded working memory, persistent memory stores, modular memory models, and deletion-aware multimodal memory, rather than as a single architecture.

## 2. Formal objectives and information-based criteria

The most explicit information-theoretic formalization is MemFly’s Memory Information Bottleneck objective,
\[
\min_{\pi}\mathcal L_{\mathrm{IB}}(M_t)=\min_{\pi}\left[I(X_{1:t};M_t)-\beta I(M_t;Y)\right],
\]
where \(X_{1:t}\) is the interaction history, \(M_t\) the memory state, and \(Y\) unknown future reasoning tasks [2602.07885]. The first term measures representational complexity or compression, and the second measures relevance to future tasks. Because direct optimization is intractable, MemFly uses structural proxies such as local coherence and global navigability, together with an online merge/link/append policy.

The paper titled InfoMem introduces a different but related objective: answer-conditioned information gain as a reward for chunk-wise memory agents,
\[
r_{\mathrm{gain}}(x,M,y^*)=\frac{1}{|y^*|}\log P_\theta(y^*\mid x,M)-\frac{1}{|y^*|}\log P_\theta(y^*\mid x,\emptyset).
\]
This quantity measures the average per-token log-likelihood gain of the ground-truth answer \(y^*\) when conditioning on the final memory \(M\) rather than a null memory [2606.03329]. The reward is applied only to successful rollouts, normalized across successful trajectories in a GRPO group, and then added to the binary outcome reward with \(\beta=0.2\). Analyses show that effective final-memory rewards should operate on successful trajectories, be normalized before reward composition, and be conditioned on the answer rather than the query [2606.03329].

MeMo offers a complementary formal perspective. It formalizes the knowledge integration problem as finding \((\Phi,f)\) such that, for all queries \(q\in Q\),
\[
\mathbb P\big[f(M_\theta,\Phi(\mathcal D),q)=a^\star(q)\big]=1,
\]
while keeping the Executive \(M_\theta\) frozen and delegating corpus-specific knowledge to a separate memory model \(M_\varphi\) with \(m\ll p\) parameters [2605.15156]. In MeMo, the memory is not a retriever over documents but a small QA model specialized to a target corpus.

These formulations are not identical, but they converge on the same principle: memory quality is defined by the extent to which retained information reduces uncertainty about future answers or tasks, not by storage volume alone. This suggests a common InfoMem criterion across architectures: retained state should be evaluated by its incremental utility for downstream inference.

## 3. Memory representations and write/update mechanisms

Recent InfoMem systems differ most sharply in what they treat as the memory object and how that object is updated. MeMo encodes corpus knowledge into a dedicated parametric Memory \(M_\varphi\) trained on a synthetic reflection QA dataset
\[
Q_{\mathrm{final}}=Q_{\mathrm{ver}}\cup Q_{\mathrm{ent}}\cup Q_{\mathrm{cross}},
\]
where \(Q_{\mathrm{ver}}\) contains self-contained fact questions, \(Q_{\mathrm{ent}}\) entity-surfacing questions, and \(Q_{\mathrm{cross}}\) cross-document synthesis questions [2605.15156]. MemFly stores a stratified graph \(\mathcal G_t=(\mathcal V_t,\mathcal E_t,\Phi_t)\) over Notes, Keywords, and Topics, and updates it online via an LLM-gated choice among Merge, Link, and Append operations [2602.07885]. Infini Memory stores topic documents with YAML-like headers and bullet-level metadata such as `seq`, `time`, and `source`, using a `CURRENT` buffer that is periodically rewritten and routed into the document library [2606.10677]. InfMem keeps a bounded textual memory \(m_t\) and updates it through PreThink–Retrieve–Write with evidence-aware joint compression [2602.02704]. AdaMem combines a standard memory store with a role-specific natural-language Memory Policy \(\mathcal P=\{C_1:\mathcal T_1,\dots,C_n:\mathcal T_n\}\), refining the policy weekly via patch-style self-reflection and failure rollback [2606.21144].

The main memory forms can be summarized as follows.

| System | Memory representation | Update mechanism |
|---|---|---|
| MeMo | Separate memory LLM \(M_\varphi\) | Offline reflection generation and supervised fine-tuning |
| MemFly | Note–Keyword–Topic graph | Gated Merge/Link/Append and topic reclustering |
| Infini Memory | Topic-structured Markdown documents | Buffering, rewrite, routing, split, merge |
| InfMem | Bounded textual memory \(m_t\) | PreThink–Retrieve–Write compression |
| AdaMem | Memory store plus role-specific policy text | Weekly reflection patch and rollback |

MeMo’s cross-document structure is pushed into training time rather than retrieval time. Its Generator first extracts direct and indirect facts, consolidates questions with common context, rewrites them to be self-contained, builds entity-surfacing questions, and finally synthesizes cross-document questions over topically related groups of chunks [2605.15156]. Appendix ablations show that removing the cross-document synthesis step collapses performance, with NarrativeQA dropping from 24.0% to 6.37% and MuSiQue from 42.9% to 24.17% [2605.15156].

MemFly’s update policy is more explicitly online. A new observation is transformed into a Note \(n_t=(r_t,c_t,\mathbf h_t,\mathcal K_t)\), where \(r_t\) is raw content, \(c_t\) a denoised context, \(\mathbf h_t\) an embedding, and \(\mathcal K_t\) a set of Keywords [2602.07885]. If redundancy exceeds \(\tau_m\), the note is merged; if complementarity exceeds \(\tau_l\), a relation edge is created; otherwise the note is appended. Topic evolution then reclusters the keyword co-occurrence graph using Leiden modularity subject to size constraints [2602.07885].

Infini Memory emphasizes maintainable text. A flush occurs when
\[
\mathrm{flush}(C)=\bigl(|C|\geq \tau_{\mathrm{tok}}\bigr)\lor\bigl(\Delta t(C)\geq \tau_{\mathrm{time}}\bigr),
\]
with \(\tau_{\mathrm{tok}}=5000\) in experiments [2606.10677]. The buffer is rewritten into a normalized draft, routed to existing or new topic documents through `SafePlanUpdate`, and then merged or split under token thresholds. Contradictions are resolved by recency, with preference for user-sourced entries over AI-sourced ones when they conflict [2606.10677].

AdaMem’s key move is to learn the write criterion itself. Incorrect weekly QA answers are analyzed as either recall failures, where relevant information was present in the dialogue but not extracted, or precision failures, where noisy memory or retrieval degraded answering. The system proposes a candidate patch \(\Delta\mathcal P_t\), applies it to obtain \(\mathcal P'_{t+1}\), and rolls back if historical validation does not improve [2606.21144].

## 4. Retrieval, read-time control, and efficiency

Read-time control is a second major axis of InfoMem design. MeMo eliminates a retrieval index at inference time: the Executive first generates clue-probing sub-queries for Memory, then performs entity identification, then answer seeking and synthesis through a purely textual multi-turn protocol [2605.15156]. Each query triggers Stage 1 with a fixed or bounded number of Memory calls, Stage 2 with up to 7 entity-identification iterations, and Stage 3 with up to 8–15 answer-seeking iterations; the resulting retrieval cost is \(O(T)\) with \(T\) fixed by protocol design rather than corpus size [2605.15156]. This is one of the clearest instantiations of the InfoMem requirement that inference cost be independent of corpus size once memory is trained.

SwiftMem instead retains retrieval but makes it query-aware and sub-linear. Its temporal index organizes each user’s episodes in sorted timelines \(\mathcal L_u\) that support range queries in \(O(\log N_{\mathrm{mem}})\), while the semantic DAG-Tag index routes a query to relevant topics with
\[
T_{\mathrm{query}}=O\bigl(k\cdot(\log|V|+D_{\max})\bigr)
\]
before dense similarity search is run on the reduced candidate set [2601.08160]. On LoCoMo, SwiftMem reports 11 ms search latency and 1289 ms total latency, compared with 522 ms search for Zep and 835 ms for Nemori; the abstract summarizes this as \(47\times\) faster search than state-of-the-art baselines while maintaining competitive accuracy [2601.08160].

MemFly uses a tri-pathway retrieval mechanism that combines macro-semantic localization over Topics, micro-symbolic anchoring over Keywords, and topological expansion over Note–Note edges, then fuses ranked lists with Reciprocal Rank Fusion [2602.07885]. An Iterative Evidence Refinement loop tests whether current evidence is sufficient, and if not, generates a refined sub-query and retrieves again up to \(I_{\max}=3\) iterations [2602.07885].

Infini Memory’s read path is explicitly agentic. Infini Memory-H selects candidate documents by summary and then uses BM25 over document partitions, whereas Infini Memory-A lets the LLM iteratively call tools such as `list_docs`, `search`, `grep_doc`, and `read_lines` for up to 7 iterations before answering [2606.10677]. If evidence is insufficient, a BM25 fallback over heading partitions is invoked, and the `CURRENT` buffer is appended so that recently acquired information is immediately available [2606.10677].

A human-facing variant of InfoMem appears in Memoro, a wearable audio memory assistant that continuously transcribes speech, stores timestamped text chunks with `all-MiniLM-L6-v2` embeddings, retrieves top-\(k=10\) semantically similar memories, and uses an LLM to provide minimal suggestions in Query Mode or Queryless Mode [2403.02135]. In a study of \(N=20\), Memoro reduced device interaction time and increased recall confidence while preserving conversational quality [2403.02135].

Taken together, these systems suggest a shift from single-step top-\(k\) retrieval toward controlled memory access: bounded multi-turn querying, topic-aware routing, semantic-symbolic-topological fusion, and explicit stopping criteria.

## 5. Maintainability, forgetting, and deletion

Maintainability has two distinct meanings in the InfoMem literature. The first is model-side stability under updates. MeMo avoids catastrophic forgetting by never training the Executive \(M_\theta\); all adaptation occurs in the separate Memory \(M_\varphi\), which preserves the base model’s general capabilities and safety alignment while allowing plug-and-play use even with closed-source systems such as Gemini-3-Flash [2605.15156].

The second is persistent-store maintenance. Infini Memory treats long-term memory as a lifecycle problem of writing, consolidating, and reading. `RewriteCurrent` deduplicates repeated entries, keeps newer duplicates when they differ only by `seq`, resolves contradictions so that newer `seq` wins, and prefers the user’s entry over AI-sourced restatements when facts conflict [2606.10677]. The default thresholds in experiments flush `CURRENT` at 5000 tokens, split topic documents at `token_count ≥ 5000`, and merge small documents when `token_count ≤ 1000` and summaries are sufficiently similar [2606.10677]. On LongMemEval, removing split and merge lowers Infini Memory-H from 76.0% to 69.3%, indicating that structural maintenance is not cosmetic but functionally important [2606.10677].

A third, more stringent notion of forgetting is behavioral erasure after a deletion request. MemLeak introduces the Information Provenance Graph \(G_f=(V_f,E_f,\delta)\), where \(\delta\) labels each representation node as `addressable`, `linked`, or `persistent` depending on its deletion affordance [2606.29788]. It defines Forgetting Completeness as
\[
FC_S(f)=\frac{|\{v\in V_f: v \text{ is deleted by } S\}|}{|V_f|},
\]
but shows empirically that storage-level deletion can still fail behaviorally [2606.29788]. Direct probing of deletion-capable text systems yields less than 1% leakage, yet retained correlated text enables 18.3% recovery and retained images enable 12.0% recovery, with a 0.0% blind baseline and 0.3% false-positive rate; 47% of image leaks are not text-recoverable [2606.29788]. Content-aware semantic deletion, implemented by a VLM auditor over retained images, reduces the image residual to 2.0%, and dual-annotator human validation reports \(\kappa=0.88\) [2606.29788].

These results complicate any purely text-centric notion of InfoMem. A memory system can be updateable, compact, and retrieval-efficient yet still fail at fact-level forgetting if residual multimodal traces remain behaviorally accessible.

## 6. Empirical landscape, design lessons, and open directions

The reward mechanism titled InfoMem demonstrates that a final-memory reward can improve chunk-wise memory agents under a fixed GRPO setup. Relative to outcome-only GRPO, InfoMem raises CorpusQA from 16.413 to 19.453, LongMemEval from 10.000 to 12.800, MRCR-8needle from 0.063 to 0.279, and RULER synthetic QA from 34.735 to 36.848 [2606.03329]. Ablations show that removing \(r_{\mathrm{gain}}\) normalization or replacing answer-conditioned gain with query-conditioned QueryPMI substantially degrades performance, and training diagnostics show that QueryPMI induces a degenerate strategy in which memory starts to repeat the query [2606.03329].

MemFly reports the best average F1 and BLEU-1 across GPT-4o-mini, GPT-4o, Qwen3-8B, and Qwen3-14B on LoCoMo, with especially strong gains in Open Domain and Adversarial categories [2602.07885]. Its ablations are notable because removing the entire gated update lowers F1 on Qwen3-8B from 38.62 to 27.97 and Recall from 62.22 to 42.11, whereas removing Link can be more harmful than removing Merge in some categories, especially Adversarial, indicating that relational structure is central to relevance rather than a secondary embellishment [2602.07885].

InfMem shows that System-2-style memory control can improve both quality and efficiency for ultra-long documents. On Qwen3-1.7B, Qwen3-4B, and Qwen2.5-7B, InfMem improves average absolute accuracy over MemAgent by +10.17, +11.84, and +8.23 points, respectively, while reducing inference time by \(3.9\times\) on average and up to \(5.1\times\) via adaptive early stopping [2602.02704]. Its early-stop shaping reward
\[
R_{\mathrm{early}}=\gamma^{d-1}
\]
aligns stopping time with the first step at which the answer is already derivable from memory [2602.02704].

MeMo shows a different trade-off: a corpus-specific memory model can be trained once and then attached to stronger Executives later. With a Qwen2.5-14B Memory and a Gemini-3-Flash Executive, MeMo reports 66.67% on BrowseComp-Plus, 53.58% on NarrativeQA, and 60.20% on MuSiQue, while remaining robust to added negative documents in the corpus used for data synthesis [2605.15156]. AdaMem addresses personalization rather than open-domain corpus integration, improving QA by up to +9.0% over the uniform Mem0 baseline while shrinking memory volume by 9% [2606.21144]. Infini Memory focuses on persistent agent memory and reaches 64.7% overall on MemoryAgentBench, with ablations showing that topic-structured maintenance and iterative evidence inspection improve complementary aspects of long-term memory use [2606.10677].

Across these systems, several design lessons recur. Separation of concerns between reasoning and memory recurs in MeMo and InfMem; write control recurs in AdaMem and Infini Memory; structural retrieval recurs in MemFly, SwiftMem, and Infini Memory; and answer-conditioned or task-conditioned utility recurs in both the reward-based InfoMem formulation and the broader information bottleneck view [2605.15156] [2606.21144] [2602.07885] [2606.03329]. A plausible implication is that future InfoMem research will increasingly optimize memory along four coupled axes at once: what to write, how to structure it, how to access it under bounded cost, and how to guarantee that both updates and deletions remain behaviorally faithful.

Source: https://www.emergentmind.com/topics/infomem