SEA-Mem: Self-Evolving Agentic Memory

• SEA-Mem is an adaptive memory mechanism that persistently stores, retrieves, and revises experiences across task boundaries.
• It employs architectural patterns like dual-process organization and forward/backward coupling to integrate memory into agent decision-making.
• SEA-Mem enhances efficiency and accuracy in various domains by evolving memory policies through feedback, reflection, and online adaptation.

Self-Evolving Agentic Memory (SEA-Mem) denotes a class of memory mechanisms for agents that persist experience across task boundaries, retrieve prior episodes or abstractions when similar contexts recur, and continually revise memory contents through feedback, reflection, or planning outcomes. In the literature, SEA-Mem appears both as an explicit module name and as a broader architectural pattern: case-based memory in PET theranostics, trajectory memory in agentic image restoration, probe-driven self-repair for conversational memory, dual-process cognitive memory for personalization, and memory-policy meta-evolution for general LLM agents (Chen et al., 14 Mar 2026, Cui et al., 27 Jun 2026, Lin et al., 19 Mar 2026, Fei et al., 8 Jun 2026). Across these instantiations, the unifying idea is that memory is not a passive retrieval store but an adaptive subsystem that stores structured experience, guides future action, and changes its own contents or usage policies over time.

1. Conceptual scope and formal definitions

A concise formalization of self-evolving agents treats the agent as

$$\mathcal{SEA} = (E, \mathcal{A}, \mathcal{M}, \Phi, \Psi),$$

where $E$ is an open-ended environment, $\mathcal{A}$ is an open, extensible action space, $\mathcal{M}$ is persistent memory, $\Phi : \tau \rightarrow \Delta\mathcal{M}$ distills a trajectory into candidate memory updates, and $$\Psi : \mathcal{M}_t \times \Delta\mathcal{M} \rightarrow \mathcal{M}_{t+1}$$ selectively integrates them (Jiang et al., 10 Apr 2026). In that formulation, memory is explicitly stratified as

$$\mathcal{M} = \mathcal{M}_p \cup \mathcal{M}_e \cup \mathcal{M}_d,$$

with procedural, episodic, and declarative components (Jiang et al., 10 Apr 2026).

A broader systems view places memory inside a general self-evolving optimization loop. The agent system $\mathcal{A}$, including its memory module, is optimized against environment-derived objectives by an optimizer over a search space of agent configurations; in this view, memory schemas, retrieval rules, summarization prompts, and forgetting policies are all part of the evolvable agent configuration (Fang et al., 10 Aug 2025). This framing is consistent with EvoMemBench, which treats memory as continuously updated state $m_t^e$ inside or across episodes and distinguishes in-episode versus cross-episode memory, and knowledge-oriented versus execution-oriented memory (Wang et al., 18 May 2026).

A recurrent misconception in this literature is that memory is adequately modeled as retrieval alone. DCPM states that “Long-term memory … is currently treated as a retrieval problem” and argues that this collapses belief revision, causal coupling, and cross-domain abstraction into a single recall surface (Fei et al., 8 Jun 2026). SEA-Mem, by contrast, is defined by persistence plus transformation: what is stored changes, and what the agent can do changes because memory is reorganized, distilled, or repaired.

2. Recurrent architectural patterns

A dominant pattern is dual-process organization. In SEAR, real-world image restoration is decomposed into a Deliberate Planner (System 2) and an Intuitive Executor (System 1): System 2 uses an LLM macro-agenda and Pruning-Aware Monte Carlo Tree Search, while System 1 retrieves a stored trajectory from episodic memory using degradation-aware fingerprints and either executes it directly or passes it to System 2 as a warm start (Cui et al., 27 Jun 2026). DCPM uses a related split: a synchronous “daytime writer” that records raw inputs, facts, identities, and supersedes chains, and an asynchronous “nighttime engine” that induces schemas, intentions, and cross-domain core schemas (Fei et al., 8 Jun 2026).

A second pattern is forward/backward coupling of the memory cycle. MemMA explicitly models construction, retrieval, and utilization as a closed loop. On the forward path, a Meta-Thinker guides memory construction and iterative retrieval; on the backward path, the system synthesizes probe QA pairs, verifies provisional memory, diagnoses failures, and turns them into repair actions before memory is finalized (Lin et al., 19 Mar 2026). This makes memory errors actionable at the memory-bank level rather than only at the policy level.

A third pattern separates raw experience from higher-level usage policy. Live-Evo maintains an Experience Bank for “what happened” and a Meta-Guideline Bank for “how to use it”; for each new task, it retrieves experiences, retrieves a meta-guideline, compiles task-adaptive guidance, and then updates memory weights and meta-guidelines from contrastive feedback (Zhang et al., 2 Feb 2026). Mem²Evolve introduces a related duality between Experience Memory and Asset Memory, where accumulated experience guides dynamic creation of tools or expert agents, and newly created assets in turn generate new experience (Cheng et al., 13 Apr 2026).

A fourth pattern internalizes memory into the policy itself. TMEM introduces fast parametric memory $\Delta_t$ so that actions are sampled from $E$0, with extraction actions generating QA-style supervision that updates LoRA weights online within the episode (Ren et al., 3 Jun 2026). This is the strongest version of SEA-Mem’s “self-evolving” property: memory does not merely alter prompt context but changes future policy dynamics during the same rollout.

3. Memory representations, indexing, and retrieval

SEA-Mem systems differ sharply in what they store and how they address it. Some store trajectories indexed by structured state fingerprints; some store symbolic case records; some maintain revision chains and schema graphs; others use atomic linked notes. The storage unit strongly constrains what can later be reused.

System	Stored unit	Retrieval basis
SEAR	$E$1	fingerprint similarity + $E$2
TheraAgent	$E$3	weighted feature matching
DCPM	facts, identities, supersedes chains, schemas	hybrid vector search + chain/graph lift
A-MEM	$E$4	cosine similarity + linked notes

In SEAR, the key is a degradation-aware fingerprint

$E$5

where $E$6 primitive degradation types and $E$7 are discrete intensities. Retrieval uses a potential score

$E$8

with $E$9 in experiments, balancing historical peak reward against fingerprint mismatch (Cui et al., 27 Jun 2026).

TheraAgent uses a more symbolic case memory. Each memory entry is a structured patient feature profile and documented outcomes, $\mathcal{A}$0, indexed by prognostic features such as PSMA level, liver/lung metastases, and prior chemotherapy. Retrieval is by weighted feature matching, and the empirical case rate over retrieved neighbors is

$\mathcal{A}$1

which is then supplied to the reasoning module as case evidence (Chen et al., 14 Mar 2026).

DCPM layers retrieval. It stores raw inputs, facts, identities, supersedes chains, and graph nodes for schemas, intentions, and core schemas. The read path uses hybrid dense cosine plus BM25 over Qdrant, traverses supersedes chains to recover belief trajectories, and then lifts matched facts into schema- and core-schema-level abstractions through evidence_vdb_ref edges (Fei et al., 8 Jun 2026).

A-MEM, by contrast, represents each memory note as

$\mathcal{A}$2

with raw content, timestamp, keywords, tags, contextual description, embedding, and links. Retrieval is by cosine similarity

$\mathcal{A}$3

followed by expansion through linked notes in the same conceptual neighborhood (Xu et al., 17 Feb 2025).

MemSkill adds another layer: the memory bank remains textual, but the operations that create and revise it become evolvable skills selected by a controller. Retrieved memories are fed to an executor together with selected skills such as “Capture Temporal Context” or “Track Object Location,” and memory updates are emitted as INSERT, UPDATE, DELETE, or NOOP actions (Zhang et al., 2 Feb 2026).

4. Mechanisms of self-evolution

Self-evolution typically combines four operations: distillation, validation, maintenance, and adaptation of the memory policy itself.

Distillation converts expensive or lengthy experience into reusable memory. In SEAR, a successful planned trajectory $\mathcal{A}$4 is distilled into memory if its reward exceeds a threshold $\mathcal{A}$5, with new entries added or existing summary statistics updated: $\mathcal{A}$6 A retrieved trajectory is accepted only if its reward exceeds a shortcut threshold $\mathcal{A}$7; otherwise it becomes a warm start for further planning (Cui et al., 27 Jun 2026).

Validation and repair are explicit in MemMA. After building provisional memory for a session, the system synthesizes a probe set

$\mathcal{A}$8

runs the normal retrieval-and-answer pipeline on those probes, identifies failures, and converts them into repair facts

$\mathcal{A}$9

which are then consolidated into the memory bank via SKIP, MERGE, or INSERT decisions (Lin et al., 19 Mar 2026).

Online adaptation is central in Live-Evo. Each experience carries a weight, and after contrastive evaluation the weight of retrieved experiences is updated by

$\mathcal{M}$0

Experiences that help are reinforced; misleading or stale ones decay. New experiences are added only for worst-case tasks and only after “Verify Before Update” confirms improvement, while failed guidance can spawn new meta-guidelines (Zhang et al., 2 Feb 2026).

Pattern induction also appears in symbolic case systems. In TheraAgent, when three or more cases share a feature combination, SEA-Mem computes subgroup outcome rates. The paper’s example is: among 8 patients with high PSMA and no liver metastasis, 7 achieved PSA response, yielding the pattern “high PSMA + no liver metastasis $\mathcal{M}$1 77.8% PSA response” (Chen et al., 14 Mar 2026).

Policy evolution is made explicit in MemSkill and MemEvolve. MemSkill treats memory operations as skills, optimizes Top-$\mathcal{M}$2 skill selection with PPO, and periodically evolves the skill bank from clustered hard cases using an LLM designer with rollback if reward degrades (Zhang et al., 2 Feb 2026). MemEvolve elevates this to the architecture level: each candidate memory system is a tuple

$\mathcal{M}$3

corresponding to encode, store, retrieve, and manage modules, and an outer-loop meta-evolution process selects and redesigns these modules based on performance, cost, and delay (Zhang et al., 21 Dec 2025).

5. Empirical evidence across tasks and domains

The evidence base is heterogeneous but consistent on one point: persistent memory can materially change efficiency, robustness, or accuracy, yet the effect is highly task- and memory-type-dependent.

SEAR provides a clear amortization result. On Group B, full SEAR achieves PSNR $\mathcal{M}$4 with Tool Calls $\mathcal{M}$5, while removing memory yields PSNR $\mathcal{M}$6 with Tool Calls $\mathcal{M}$7. On the same dataset, memory routing reports 98% memory hit and 62% accepted shortcuts (Cui et al., 27 Jun 2026). This is a canonical SEA-Mem effect: similar end quality, much lower planning cost.

TheraAgent reports that full performance on real clinical data is 75.7% overall accuracy, versus 70.0% without SEA-Mem; on synthetic data it reaches 87.0%, and on misleading synthetic cases 78.8% (Chen et al., 14 Mar 2026). These results indicate that case-based memory adds measurable value even when data are scarce.

MemMA shows that memory self-repair affects downstream QA materially. On LoCoMo with the Single-Agent backend and GPT-4o-mini, MemMA$\mathcal{M}$8 reaches 84.87 ACC, while removing self-evolution drops ACC to 73.68, and removing iterative retrieval drops it to 70.39 (Lin et al., 19 Mar 2026). DCPM shows a complementary pattern: System 2 contributes little on surface recall but substantially more on implicit cross-session inference, contributing up to +5.20 on PersonaMem-v2 and much less on LongMemEval, which matches its emphasis on cross-domain abstraction rather than span recall (Fei et al., 8 Jun 2026).

Mem²Evolve reports average improvement of 18.53% over standard LLMs, 11.80% over agents evolving solely through experience, and 6.46% over those evolving solely through asset creation; removing Tool Creation drops average performance by 10.28, while removing Agent Memory drops it by 4.73 (Cheng et al., 13 Apr 2026). Live-Evo, in a genuinely online setting, improves Brier score by 20.8% and market returns by 12.9% on Prophet Arena (Zhang et al., 2 Feb 2026). AdaMEM extends the same theme to test-time adaptive retrieval and strategy synthesis, with relative gains of up to 13% on ALFWorld and 11% on WebShop (Zhang et al., 4 Jun 2026).

Benchmarking work tempers these successes. EvoMemBench finds that no single memory form works consistently across all settings, that long-context baselines remain highly competitive, and that memory helps most when current context is insufficient or tasks are difficult (Wang et al., 18 May 2026). SEA-Eval shows that under identical success rates, token consumption can differ by up to $\mathcal{M}$9, arguing that success rate alone can mask pseudo-evolution (Jiang et al., 10 Apr 2026). MemoryArena further shows that agents with near-saturated performance on long-context recall benchmarks perform poorly in interdependent multi-session tasks, exposing a gap between recall-style evaluation and genuinely agentic memory use (He et al., 18 Feb 2026).

Taken together, these results suggest that SEA-Mem is most effective when it captures structure that the base context window or a static retriever cannot reliably reconstruct: trajectory reuse, recurrent case patterns, belief evolution, or domain-specific workflows.

6. Limitations, failure modes, and open problems

Current SEA-Mem systems remain sensitive to representation quality, retrieval reliability, and evaluation protocol. SEAR depends on MLLM-based degradation diagnosis and on the coverage of the tool library; cold-start overhead remains high until enough trajectories accumulate (Cui et al., 27 Jun 2026). TheraAgent is built on only 35 real patients plus 400 synthetic cases, lacks explicit pruning or learned similarity metrics, and raises privacy concerns because case memory is institution-specific (Chen et al., 14 Mar 2026).

MemMA is evaluated only on dialogue-centric memory tasks, assumes meaningful session boundaries, and incurs nontrivial LLM cost because construction guidance, iterative retrieval, probe generation, repair, and consolidation all require additional calls (Lin et al., 19 Mar 2026). DCPM is sensitive to threshold choice, can over-fire supersedes chains on transient states, and stores schemas in natural language, which limits algebraic manipulation and complicates structured revision (Fei et al., 8 Jun 2026). Live-Evo assumes dense feedback and uses a conservative experience-admission rule, which can delay adoption of subtle but valuable heuristics (Zhang et al., 2 Feb 2026). Mem²Evolve has no explicit pruning or de-duplication policy and depends on a sandboxed execution environment for generated tools (Cheng et al., 13 Apr 2026).

At the system level, the survey literature identifies broader issues: benchmark saturation, metric misalignment, backbone dependence, and the latency/throughput overhead of memory maintenance (Jiang et al., 22 Feb 2026). SEA-Eval’s emphasis on token convergence and MemoryArena’s POMDP framing both point to a deeper open problem: memory quality should be evaluated by whether it supports stable belief-state tracking and lowers future reasoning cost, not only by whether it retrieves locally relevant text (Jiang et al., 10 Apr 2026, He et al., 18 Feb 2026).

A plausible implication is that mature SEA-Mem systems will need hybridization rather than monolithic design: lightweight semantic buffers for cheap access, structured or graph-based stores for precise relational state, procedural or trajectory memory for repeated workflows, and meta-level controllers that decide when each layer should be updated, queried, compressed, or ignored. That implication is consistent with EvoMemBench’s finding that no single memory form is universally dominant, and with the broader survey view that self-evolving agents should optimize memory as one component of a larger lifelong agentic system (Wang et al., 18 May 2026, Fang et al., 10 Aug 2025).