Stateless vs. Stateful Agent Memory: The Enterprise Architecture Divide

A new preprint — "Stateless Decision Memory for Enterprise AI Agents" by Vasundra Srinivasan — argues that mutable agent memory is an operational liability in regulated environments. This video breaks down Deterministic Projection Memory (DPM), a framework that replaces continuous memory updates with an append-only event log and a single late-bound projection call at decision time. The result: dramatically smaller audit surfaces, deterministic replay, and multi-tenant isolation by default.

Key topics covered in order:

• Why enterprise adoption favors stateless retrieval — not because of answer quality, but because immutable indexes, stateless queries, and pure functions already satisfy regulatory constraints that mutable memory systems must retrofit.

• How DPM works — an append-only event log plus a single projection function that materializes a decision-ready memory view (facts, reasoning, compliance notes) in one LLM call. No intermediate state mutations, no consolidation passes.

• The empirical results on LongHorizon-Bench — at tight memory budgets (≈20× compression), DPM achieves +0.52 factual precision and +0.53 reasoning coherence over incremental summarization, while requiring only 2 LLM calls versus 83–97 for the stateful baseline.

• The industry counter-case — Mem0, Zep, Letta, and Supermemory all treat persistent memory as a first-class architectural layer. Their strongest arguments center on personalization, temporal reasoning, and agent identity. Bilt's million-agent deployment on Letta is a concrete production proof point.

• The real fault line — this is not pro-memory vs. anti-memory. It is two incompatible definitions of what memory is for. Stateful memory wins where the product is personalization and continuity. Stateless projection wins where the product is a governed, auditable decision with legal replay requirements.

The paper: "Stateless Decision Memory for Enterprise AI Agents" — arxiv.org/abs/2604.20158
Companion paper: "Four-Axis Decision Alignment for Long-Horizon Enterprise AI Agents"

#agentmemory #llm #enterpriseai #rag #agents #memoryarchitecture #aiengineering #systemdesign

📑 Chapters:
0:00 The enterprise case against mutable agent memory
0:30 Why this matters: the stateless vs. stateful divide
1:07 How DPM works: append-only logs and late-bound projection
1:50 Why enterprise RAG already had the right architectural shape
2:22 The projection function: one call replaces the entire update loop
2:55 Audit surface cost: 2 LLM calls vs. 83–97
3:25 Budget-conditional results on LongHorizon-Bench
3:55 Determinism experiment: replay hashes and prefix stability
4:20 The industry counter-case: Mem0, Zep, Letta, Supermemory
5:05 Sorting by workload: personalization vs. regulated decisions
5:50 The rule: when agents decide, memory becomes evidence

#agent memory #deterministic projection memory #DPM #enterprise AI #stateless retrieval #stateful agents #RAG #Mem0 #Letta #Zep #Supermemory #audit surface #LLM architecture #event sourcing #context engineering

Видео Stateless vs. Stateful Agent Memory: The Enterprise Architecture Divide канала The Bearded AI Guy