MemCollab: Cross-Agent Memory Framework

• MemCollab is a cross-agent memory framework that aggregates shared solution patterns from diverse LLM-based agents to improve reasoning across tasks.
• It employs contrastive trajectory distillation to extract transferable invariants while filtering out agent-specific biases and heuristics.
• Empirical results demonstrate significant gains in accuracy and efficiency on benchmarks like MATH500 and HumanEval through task-aware memory retrieval.

MemCollab is a cross-agent, collaborative memory framework that enables heterogeneous LLM-based agents to construct, share, and leverage a unified agent-agnostic memory bank, thereby enhancing accuracy and efficiency across mathematical reasoning and code generation tasks. Unlike per-agent memory approaches—which conflate agent-specific heuristics with problem-solving knowledge—MemCollab employs contrastive trajectory distillation to distill shared reasoning invariants that generalize across model families, architectures, and reasoning styles. This architecture supports task-category–conditioned retrieval and serves as a foundation for collective intelligence over distributed LLM agent deployments (Chang et al., 24 Mar 2026).

1. Motivation: Cross-Agent Memory Reuse and Its Limitations

Traditional LLM agent deployments construct and store memory independently per agent, encoding not only shared solution strategies but also model-specific preferences, heuristics, and idiosyncratic tool usage. Given a set of agents $\mathcal{A} = \{A_1, ..., A_n\}$, each presented with task $x$, the generated trajectory $\tau_i^{(x)} = (r_1, ..., r_{T_i})$ entwines global task constraints with agent-local peculiarities. Naïvely transferring memory $M_1$ from agent $A_1$ to agent $A_2$ has been shown to degrade downstream performance, as agent-specific bias $b_1$ can conflict with $A_2$'s reasoning style. Empirical results demonstrate accuracy drops in Qwen-7B on MATH500 when using Qwen-32B-derived memory, underscoring the need for agent-agnostic memory that encodes shared invariants $s$ and excludes agent-unique elements $b_i$ (Chang et al., 24 Mar 2026).

2. Formalization: Problem Setup and Agent-Agnostic Memory Construction

MemCollab formalizes the construction of a global memory bank $x$0 that aggregates knowledge from all available agent trajectories on the training set:

$x$1

The goal is to maximize task accuracy and efficiency for any agent $x$2 on new task $x$3 by retrieving a relevant memory slice:

$x$4

Agent-specific memory $x$5 is formed by distillation solely from $x$6, whereas ideal agent-agnostic memory $x$7 captures only shared solution patterns, maximizing transferability.

3. Methodology: Contrastive Trajectory Distillation

The core methodological innovation of MemCollab is to extract abstract, agent-agnostic constraints via contrastive distillation:

• Agents $x$8 (weaker) and $x$9 (stronger) solve training instances $\tau_i^{(x)} = (r_1, ..., r_{T_i})$0, producing trajectories $\tau_i^{(x)} = (r_1, ..., r_{T_i})$1.
• An indicator function $\tau_i^{(x)} = (r_1, ..., r_{T_i})$2 identifies the preferred (correct) trajectory $\tau_i^{(x)} = (r_1, ..., r_{T_i})$3 and the less-preferred (incorrect) $\tau_i^{(x)} = (r_1, ..., r_{T_i})$4.
• A backbone LLM is prompted to contrast $\tau_i^{(x)} = (r_1, ..., r_{T_i})$5, extracting pairs of (\emph{violated}, \emph{invariant}) reasoning patterns $\tau_i^{(x)} = (r_1, ..., r_{T_i})$6:

$\tau_i^{(x)} = (r_1, ..., r_{T_i})$7

• Each pair is converted to an abstract memory entry $\tau_i^{(x)} = (r_1, ..., r_{T_i})$8“enforce $\tau_i^{(x)} = (r_1, ..., r_{T_i})$9”; “avoid $M_1$0”$M_1$1, labeled with the originating task category, and stored in $M_1$2.
• No gradient-based training is used; memory extraction and retrieval leverage prompt-based summarization and classification.

This contrastive extraction is interpretable as maximizing a proxy InfoNCE objective over positive/negative trajectory pairs:

$M_1$3

where $M_1$4 is a trajectory-memory scoring function (Chang et al., 24 Mar 2026).

4. Task-Aware Memory Retrieval and Inference

At inference, a task-aware retrieval pipeline conditions memory access on relevant categories:

• The inference agent $M_1$5 first uses its backbone LLM to predict the tuple $M_1$6—the coarse and fine category of the query $M_1$7 (e.g., “Algebra”, “Quadratic Equations”).
• Category filtering restricts retrieval to

$M_1$8

• Relevance ranking is performed via TF–IDF or embedding similarity between $M_1$9 and $A_1$0.
• The top $A_1$1 memory entries (default $A_1$2) are selected.
• These constraints serve as soft prompts to $A_1$3 during solution generation, enforcing valuable invariants and discouraging known failure paths; concrete output $A_1$4.

This architecture ensures that only category-applicable, agent-neutral constraints participate in inference-time guidance, balancing informativeness and noise (Chang et al., 24 Mar 2026).

5. Empirical Evaluation: Benchmarks, Accuracy, and Efficiency

Empirical studies validate MemCollab's improvements on mathematical and code-generation benchmarks. Representative datasets include MATH500, GSM8K, MBPP, and HumanEval. Primary metrics are exact match accuracy (math) and pass@1 (code).

Backbone	Vanilla	Self-Contrast	MemCollab	Δ over Vanilla
Qwen-7B	57.1	67.5	71.6	+14.5%
Qwen-32B	70.8	77.4	79.6	+8.8%
LLaMA-8B	41.7	—	53.9	+12.2%

Across Qwen-7B, MemCollab reduces average reasoning turns per problem: on MBPP from 3.1 to 1.4, on HumanEval from 3.3 to 1.5. Ablation demonstrates that performance increases up to $A_1$5 retrieved constraints, then declines due to noise. The two-stage retrieval (classification + similarity ranking) outperforms prompt selection or single-stage retrieval (Chang et al., 24 Mar 2026).

6. Design Insights, Limitations, and Future Directions

MemCollab’s contrastive methodology successfully isolates transferable solution invariants and removes agent-level artifacts. It distills concise, abstract constraints and prunes known failure modes, thereby serving as a generalizable reasoning resource. Limitations include:

• Over-retrieval introduces irrelevant or weakly correlated constraints, degrading accuracy.
• Memory quality depends on the fidelity of trajectory discrimination and summarization prompts.
• The cap on $A_1$6 memory entries per task may under-summarize highly complex cases.

Future avenues include scaling to larger and dynamically composed agent pools, integrating more sophisticated re-ranking of memory entries, jointly optimizing retrieval/summarization modules, and annotating memory with richer tool or solution structure. Extensions to domains with fine-grained tool use and non-trivial task taxonomies are suggested as fertile ground for MemCollab (Chang et al., 24 Mar 2026).

7. Broader Context in Multi-Agent Memory Collaboration

MemCollab represents one approach in a spectrum of multi-agent collaborative memory work. Complementary lines include reinforcement-learning–driven inter-agent optimization (Mao et al., 13 Mar 2026), mesh memory protocols for long-term multi-session agent state sharing (Xu, 21 Apr 2026), and multimodal/robotic memory frameworks employing memoization and geometric plan cache search (Wang et al., 28 Jan 2026). In contrast, MemCollab’s core contribution is its contrastive extraction of agent-agnostic knowledge, which empirically demonstrates strong gains in cross-agent transfer, correctness, and computational efficiency in heterogeneous LLM deployments.