Episodic-to-Semantic Consolidation Without Identity Drift

🕒 Published (v1): 2026-07-02 10:18 UTC · Source: Arxiv · link

Why this paper was selected

Episodic-to-semantic memory consolidation without identity drift for long-running agents

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Long-running agents deployed in regulated environments face a conflict: adapting from experience mutates the artifact whose cryptographic hash defines the agent's certified identity. This paper resolves the tension by making consolidation a deterministic function \(f: \mathcal{M}^{ep}_A \to \mathcal{M}^{sem}_A\) that writes only to a derived semantic layer excluded from the identity manifest's hash inputs, proving by construction that \(h = \text{SHA256}(M)\) remains byte-equal across arbitrary consolidation passes. Empirical validation on a 1000-decision synthetic benchmark shows a mean 79.8% reduction in unproductive planner attempts versus a calibrated Bayesian-shrunk baseline.

Problem

Conventional memory consolidation for adaptive agents (fine-tuning, prompt rewriting, reflection, policy distillation) binds the learned output to the artifact whose hash defines the agent's cryptographic identity certificate. In regulated autonomic deployments (hospitals, warehouse logistics), every consolidation pass thereby becomes a re-certification event — an operationally untenable continuous compliance burden. No existing continual-learning method (EWC, GEM, replay, reflection) preserves the cryptographic certificate while still allowing the agent to accumulate operational knowledge from lived experience.

Method

The framework models agent state as a tuple \((M, \mathcal{M}^{ep}_A, \mathcal{M}^{sem}_A, \pi)\) where: - \(M\) is a frozen identity manifest (fixed-field dataclass); identity hash \(h = \text{SHA256}(M)\) - \(\mathcal{M}^{ep}_A\) is an append-only episodic event log partitioned by \(h\) - \(\mathcal{M}^{sem}_A\) is a derived semantic knowledge table written only by the consolidation function

Structural identity invariance (Lemma 1): The manifest's hash-input field set is fixed at type-design time and explicitly excludes \(\mathcal{M}^{sem}_A\). Any \(f \in \mathcal{F}_{sem}\) (operations whose only write target is the semantic store) cannot reach \(M\)'s canonical serialization; therefore \(h\) is byte-equal before and after any such \(f\), by construction rather than runtime assertion. Corollary 1 extends this to arbitrary finite sequences of consolidation passes.

v1 deterministic consolidation algorithm: A SQL-style aggregation pass groups episodic rows by (skill, object, context), counts outcomes, computes a confidence-weighted success rate (with Bayesian shrinkage toward a prior), and upserts a single semantic fact row carrying (recommended_value, confidence, n_obs, provenance → ep_evt_id). The pass is idempotent and order-invariant: same episodic state + same rule version → byte-equal semantic snapshot.

Read-only planner interface: A downstream planner queries \(\mathcal{M}^{sem}_A\) via a read-only HTTP endpoint that surfaces top-ranked facts as grounding for the planner's intent; the planner has no write access to \(\mathcal{M}^{sem}_A\) or \(M\).

An LLM-assisted v2 (non-deterministic extraction) is deferred to future work because non-determinism would complicate audit.

Key Contributions

• Structural identity-invariance proof (Lemma 1 + Corollary 1): byte-equality of \(h\) across consolidation holds by input-set inspection of the manifest dataclass, not by runtime assertion
• Deterministic, auditable v1 consolidation algorithm: idempotent, order-invariant, produces database rows with explicit confidence, observation counts, and supporting-event provenance (full traceability from planner-consulted fact back to raw episodic events)
• Read-only planner integration interface: exposes \(\mathcal{M}^{sem}_A\) as information, not authority — no write access from the planner side
• Empirical validation: 79.8% reduction in unproductive planner attempts against a calibrated Bayesian-shrunk control, with SHA-256 identity verified byte-equal across all consolidation passes in the V2 fixture

Results

• Primary metric: mean 79.82% reduction in unproductive planner attempts (95% BCa CI [78.02%, 81.49%]) over 10 seeds on a 1000-decision synthetic benchmark
• Baseline: calibrated Bayesian-shrunk confidence control (not a naïve no-memory straw man)
• Identity verification: SHA-256 hash recomputed before and after every consolidation pass; byte-equal in all experimental runs (V2 fixture)
• Per-field correctness of consolidated facts also validated (exact numbers not quoted in the excerpt beyond the CI above)

Limitations

• Empirical validation is entirely synthetic (1000-decision benchmark); no real-world embodied robot deployment reported
• v1 algorithm handles only structured, schema-conformant episodic payloads; free-text or unstructured experience cannot be consolidated without the deferred LLM-assisted v2
• LLM-assisted v2 is explicitly out of scope because non-determinism breaks the audit chain — a significant capability gap for agents whose experience is linguistically rich
• Identity invariance holds by construction only if the implementation is well-typed and schema-correct; a typing or schema bug would require a test failure to detect
• The paper assumes the existence of an underlying autonomic-governance substrate (manifest hash + append-only episodic store) without specifying how that substrate is realized or certified in practice

Relevance to Agentic AI / LLM Agents

This work directly addresses a deployment-readiness gap for long-running LLM agents in regulated settings: all major memory-consolidation approaches (Reflexion-style reflection, skill libraries, hierarchical memory) bind the learned output to the deployed artifact, making continuous adaptation incompatible with stable certified identity. The episodic-to-semantic separation is a concrete architectural discipline that enables agents to accumulate operational knowledge without triggering re-certification, which is critical for agentic systems in safety-critical or compliance-bound domains. The read-only planner interface pattern — where consolidated knowledge is surfaced as grounded facts but the planner has no write authority — is directly applicable to tool-calling and RAG-augmented agent designs. The formal treatment of identity, audit provenance, and deterministic consolidation provides a principled foundation for trustworthy long-horizon agentic deployment.