Skip to content

Memory Depth, Not Memory Access: Selective Parametric Consolidation for Long-Running Language Agents

🕒 Published (v1): 2026-06-25 09:42 UTC · Source: Arxiv · link

Why this paper was selected

Distinguishes memory access from parametric consolidation; novel framing for long-running agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Retrieval-augmented generation (RAG) answers "what can be fetched?" but not "what should durably shape behavior after context is unloaded?" This paper formalizes that second property as memory depth, introduces a controlled evaluation protocol (loop-drift) that keeps the retrieval index intact while clearing working context, and proposes EVAF — a surprise-times-valence gated LoRA consolidation mechanism — that outperforms RAG on post-unload goal persistence while requiring only 2–3 parametric writes per 200 events.

Problem

Long-running agents rely almost exclusively on retrieval for memory, but retrieval requires the relevant text to be reinserted into context at query time. When working context is unloaded (e.g., across sessions or after long interference), retrieved facts are available yet goal-conditioned behavior collapses. No existing benchmark isolates this: public benchmarks such as LongMemEval and LoCoMo preserve retrieval access but do not probe post-unload behavioral persistence as a separate measurement axis. The gap is a missing mechanism for durable parametric tendencies that survive context unload without being reinserted as text.

Method

Loop-drift protocol. Each synthetic stream has 10 users × 200 events drawn from stable goal/preference reminders, off-topic distractors, transient opposite requests, conflicts, sibling-user contamination, and scheduled factual notes. Four probe layers: short-fact (recent explicit recall), long-fact (older noisy recall), goal persistence (goal shaping behavior under interference), and post-unload recovery (goal probe immediately after context clear, retrieval index still intact). The protocol is controlled by design; retrieval is durable across the unload event.

EVAF gate. For event \(x_t\), a write-admission score is computed as: $\(g_t = \sigma(k_s(s_t - \tau_s)) \cdot \sigma(k_v(v_t - \tau_v))\)$ where \(s_t\) is token NLL-based surprise and \(v_t\) is embedding similarity to the user's durable goal. If \(g_t > \tau_w\), the event enters a small buffer. When the buffer fills, a LoRA adapter is updated on buffered events plus replay from prior consolidations, with an L2 anchor to bound drift. Fixed constants: \(k_s=1\), \(k_v=10\), \(\tau_v=0.5\), \(\tau_w=0.5\); per-user median warmup for \(\tau_s\).

Actuation controls. Fixed-inner controllers vary only inner LoRA steps (fixed-1/2/3) with the same EVAF selection gate, separating selection from write strength. A behavior-margin controller (proof-of-concept) writes until a held-in calibration-pair margin shifts by a target amount under KL+L2 caps. Routed EVAF+RAG routes factual probes to retrieval and goal probes to EVAF to test complementarity.

Key Contributions

  • Formal distinction between memory access (retrieval) and memory depth (durable post-unload goal-conditioned behavior) for long-running agents.
  • Loop-drift protocol: a controlled stress test isolating post-unload goal persistence while preserving retrieval index integrity.
  • EVAF mechanism: surprise × valence gating into a small LoRA adapter with replay and L2 anchor; achieves high goal persistence at ~2–3 writes per 200 events.
  • Factorization of selective consolidation into two separable, coupled control dimensions: selection (gate) and actuation (write strength), with asymmetric online feedback demonstrated across GPT-2, TinyLlama, and Mistral-7B.
  • Memora boundary diagnostic: shows that append-only EVAF does not solve stale-memory deletion/update validity (non-significant, \(p=0.57\)), scoping future work toward validity-gated reconsolidation.

Results

  • Depth flip (Table 2): RAG dominates short-fact accuracy (0.956–0.973); EVAF dominates goal persistence and post-unload recovery.
  • TinyLlama: EVAF goal 0.833, post-unload 0.812 vs. RAG goal 0.396, post-unload 0.394.
  • GPT-2: EVAF goal 0.904, post-unload 0.900 vs. RAG goal 0.398, post-unload 0.394.
  • EVAF achieves this with 2.4–2.6 writes vs. Naive-LoRA's 200 writes per stream.
  • Naive-LoRA at 7B is actively harmful: Mistral-7B Naive-LoRA collapses to goal persistence 0.333 ± 0.047 (below 0.500 chance), while EVAF under miscalibrated 5-step actuation still reaches 0.362 ± 0.092.
  • Matched-gate ablation (GPT-2): EVAF mean goal/post-unload 0.790/0.763 vs. random matched gate 0.590/0.619, proving selection is not reducible to sparsity.
  • Actuation audit (Table 3): On Mistral-7B, reducing inner steps from 5 to 2 recovers goal/post 0.796 ± 0.035 / 0.775 ± 0.041 (from 0.354/0.306); Fixed-1 reaches 0.919/0.938 but saturates sibling contamination at 1.000 across all seeds.
  • Memora boundary: EVAF improves forgetting-absence from 91/222 to 95/222; McNemar \(p=0.57\) — not significant.

Limitations

  • Main protocol is fully synthetic; real-world validity is untested beyond the non-significant Memora diagnostic.
  • Heldout paraphrase generalization is weak: GPT-2 is not positive; TinyLlama is directional only — the paper claims held-in goal-conditioned behavior, not broad semantic generalization.
  • Stale-memory deletion/update is unsolved; negative-CE anti-training was unstable; reconsolidation or validity gating is identified as future work.
  • Actuation is model-dependent (no universal optimum across GPT-2, TinyLlama, Mistral-7B); high-actuation regimes increase sibling contamination.
  • 7B result establishes actuation factorization but not a complete large-model memory system; Margin controller has high seed variance on Mistral-7B.
  • Experiments use GPT-2 and TinyLlama as primary models; scale generalization to production-sized agents is not demonstrated.

Relevance to Agentic AI / LLM Agents

This paper directly addresses a critical gap in long-running agent design: the distinction between retrievable memory and durably embedded behavioral tendencies, framed as complementary learning systems (CLS) theory applied to language agents. For practitioners building persistent agents (personal assistants, continual task agents), it provides a concrete mechanism — LoRA-based selective consolidation — and a controlled evaluation protocol that can be adopted to measure post-session behavioral continuity independently of retrieval performance. The factorization of consolidation into selection × actuation with asymmetric feedback is a practically actionable insight: tuning write strength is a separable hyperparameter that must be model-calibrated, not a global constant. The Memora boundary result honestly scopes where parametric consolidation fails (memory invalidation), pointing the field toward reconsolidation and validity gating as the next required primitive for robust long-running agents.