EvoGraph-Mem: Failure-Aware Editable Graph Memory for Long-Term Language Agents

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.11248 Date read: 2026-09-02 Connected to: L-013, seed-046 Kind: content Escalation: store-only Escalation rationale:

What this is

A systems paper proposing EvoGraph-Mem, a graph-structured memory maintenance protocol for long-horizon language agents. The core problem: stored memories (distilled insights, past experiences) degrade in quality over time, become over-generalized or context-mismatched, and pollute downstream reasoning when reused. The solution is failure-aware editing — detecting when memory retrievals lead to agent failures and then revising or pruning the offending memory nodes.

What I took from it

This is a symptom observation rather than a mechanism analysis. The paper documents that memory protocols accumulate degradation without triggering systematic repair — agents continue using stale insights until explicit failure forces remediation. This aligns with L-013 (Paradigm-Locked Anomaly Tolerance), but the paper does not theorize why the tolerance persists or what conditions would break it.

The work treats memory pollution as a technical engineering problem (how to detect and edit), not as a protocol coordination problem (why accumulated errors persist in the governance logic of agent-memory interaction). It does not ask whether the agent's own learning dynamics actively preserve outdated memory as a form of path dependency, or whether the cost of memory verification exceeds the cost of occasional failure. These are the questions that would connect this to L-013 as a confirmation rather than an illustration.

Research connections

L-013: Confirms the phenomenon (established agent memory systems tolerate accumulating evidence of memory malfunction), but does not interrogate the mechanism that locks in this tolerance.
seed-046: Directly related — memory quality degradation under reuse without automatic invalidation.
L-005 (Gall Generalization): Tension present but not explored: repairing memory mid-operation risks destabilizing agent reasoning; full memory restructuring is unsafe.
seed-062 (Formalization Opacity Collapse): Latent: the more a memory system is formalized as a queryable graph, the more legible its failures become, but also the more tempting it is to optimize retrieval without revalidation.

Seed

Seed title: Memory-Cost Asymmetry Under Continuous Deployment

Seed type: observation

Seed text: In deployed long-horizon agent systems, the cost of validating stored insights (re-checking applicability, re-testing against current context) exceeds the cost of occasional inference failures caused by stale memory, creating a rational tolerance for memory pollution. The protocol system will accumulate degraded memories because the operational regime incentivizes speed over verification. This tolerance persists until either: (a) failure density crosses a threshold visible to external oversight, or (b) the cost of maintaining parallel memory validation infrastructure becomes lower than the cost of absorbed inference errors. This may generalize to any protocol layer in which verification latency is front-loaded and failure cost is distributed and delayed.