Andrii Bidochko

• Updated: July 22, 2026
• 6 min read

Shared Selective Persistent Memory for Agentic LLM Systems – An In‑Depth Overview

Direct Answer

The paper introduces Shared Selective Persistent Memory (SSPM), an architecture that lets agentic LLM systems retain only the reusable pieces of context—task specs, data schemas, tool configurations, and output constraints—while discarding transient reasoning traces. By sharing this curated memory across users and workspaces, SSPM cuts token waste, accelerates recurring tasks, and improves completion rates for complex, multi‑turn tool‑use workflows.

Background: Why This Problem Is Hard

Agentic LLMs have become the de‑facto engine for automating code generation, data extraction, and report synthesis. In practice, each interaction follows a multi‑turn “think‑plan‑execute” loop where the model calls external tools (SQL, REST APIs, version control) and refines its output. The core difficulty lies in the context bottleneck:

• Zero‑state sessions: Every new session starts with an empty prompt, forcing the model to re‑specify domain constraints, data schemas, and tool credentials that were already established in prior runs.
• Token budget limits: LLM providers charge per token. Injecting full conversation histories quickly exhausts the budget, especially when raw data tables or lengthy code snippets are involved.
• Stale reasoning noise: Past reasoning steps often contain dead‑ends, trial‑and‑error, or user corrections that are irrelevant to the next iteration, yet they occupy valuable context space.

Current workarounds either persist the entire transcript (which inflates token usage) or discard everything (which forces redundant re‑specification). Neither approach scales to enterprise workloads where dozens of agents collaborate on shared artifacts such as dashboards, data‑driven documents, or CI pipelines.

What the Researchers Propose

The authors present a four‑layer memory model that isolates “persistent” from “ephemeral” context. The persistent layer stores only the elements that are demonstrably reusable across sessions:

• Task specifications: High‑level goals, success criteria, and business rules that define what the agent should accomplish.
• Data schemas: Structural definitions of tables, JSON payloads, or API contracts that the agent must respect.
• Tool configurations: Authentication tokens, endpoint URLs, and parameter defaults for external services.
• Output constraints: Formatting rules, version‑control policies, and compliance checks that shape the final artifact.

All other information—step‑by‑step reasoning, intermediate variable values, and user corrections—remains in a short‑lived “session buffer” that is cleared after each task completes. Crucially, the persistent memory is shared across workspaces via role‑based access control, enabling multiple users or agents to read and write the same reusable context without duplication.

How It Works in Practice

Conceptual Workflow

1. Workspace initialization: When a new project is created, the system provisions a SSPM store and assigns roles (owner, contributor, viewer).
2. Context extraction: After the first successful run, the LLM tags each piece of information that belongs to one of the four persistent categories and writes it to the shared store.
3. Selective retrieval: On subsequent runs, the orchestrator prepends only the persisted items to the prompt, omitting the full transcript.
4. Zero‑token data refresh: If the underlying data source (e.g., a CSV file) changes, the system signals the agent to re‑fetch the data at runtime without re‑injecting the raw rows into the prompt.
5. Collaboration loop: Multiple agents can read the same SSPM entries, augment them with new tool configurations, or propose updated output constraints, all while respecting access policies.

Component Interaction Diagram

The architecture consists of four tightly coupled modules:

• LLM Orchestrator: Handles prompt assembly, invokes the language model, and parses the response for persistent tags.
• Selective Memory Service (SMS): A key‑value store that enforces category separation and role‑based permissions.
• Tool Adapter Layer: Normalizes calls to heterogeneous back‑ends (SQL, REST, MCP) and reports execution metadata back to the orchestrator.
• Refresh Engine: Monitors external data sources and triggers a “zero‑token” refresh flag when changes are detected.

What sets SSPM apart from naïve full‑history persistence is the disciplined pruning of context. By design, the orchestrator never re‑injects stale reasoning traces, eliminating the bias that previously caused agents to repeat past mistakes.

Evaluation & Results

Testbed and Scenarios

The researchers deployed SSPM in a production‑grade collaborative workspace platform that supports git‑versioned artifacts. They evaluated three enterprise‑scale use cases:

1. Financial reporting dashboard: Agents ingest CSVs, run SQL aggregations, and generate PowerBI‑compatible JSON.
2. Customer‑support knowledge base: Agents pull from a REST ticketing API, summarize resolutions, and update a Confluence page.
3. IoT device health monitor: Agents query an MCP server, compute anomaly scores, and push alerts to Slack.

Key Findings

• Task completion: SSPM achieved a 96 % success rate, compared with 79 % for a baseline that used no memory and 71 % for a baseline that persisted the full conversation.
• Token efficiency: Summary‑driven prompts reduced per‑invocation token consumption by a factor of 97× relative to raw data injection.
• Runtime savings: The zero‑token refresh mechanism eliminated the need to re‑invoke the LLM for recurring updates, cutting overall task time by 14×.
• Generalizability: A replication on four public datasets (including OpenML and Kaggle) showed successful zero‑token refresh in all 12 trials, confirming that the approach is not domain‑specific.
• Negative control: Full‑history persistence actively degraded performance, confirming that stale context can mislead the model.

Why This Matters for AI Systems and Agents

For AI practitioners, SSPM offers a concrete blueprint to overcome the “context explosion” problem that has limited the scalability of LLM‑driven agents. The practical benefits include:

• Reduced operating costs: Fewer tokens per request translate directly into lower cloud‑provider bills, a critical factor for enterprises running thousands of daily agent calls.
• Higher reliability: By shielding agents from irrelevant historical noise, SSPM improves deterministic behavior, making it easier to write automated tests and SLAs.
• Collaborative reuse: Shared memory enables teams to build a common knowledge base of tool configurations and data contracts, accelerating onboarding and cross‑project consistency.
• Compliance friendliness: Role‑based access to persistent memory aligns with data‑governance policies, allowing auditors to trace exactly which configuration was used for a given output.

These advantages map directly onto the capabilities of modern AI platforms. For example, the Enterprise AI platform by UBOS already provides a plug‑in architecture that can host a Selective Memory Service, making it straightforward to adopt SSPM without rebuilding the entire stack.

What Comes Next

While the results are compelling, several open challenges remain:

• Dynamic category expansion: Current work fixes four categories; future research could explore adaptive schemas that learn new reusable patterns on the fly.
• Fine‑grained provenance: Tracking which agent contributed each persistent entry would enable richer audit trails and conflict resolution in multi‑owner workspaces.
• Cross‑model interoperability: Extending SSPM to support heterogeneous LLM back‑ends (e.g., Claude, Gemini) would broaden its applicability.
• Security hardening: Integrating secret‑management solutions (e.g., HashiCorp Vault) with the Tool Configuration store could further reduce credential leakage risk.

Addressing these gaps will likely involve tighter integration with orchestration frameworks and deeper research into memory‑aware prompting techniques. Organizations interested in pioneering these extensions can join the UBOS partner program, which offers early‑access APIs, co‑development resources, and joint‑marketing opportunities.

References

For the full technical exposition, see the original arXiv paper titled “Shared Selective Persistent Memory for Agentic LLM Systems.”

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.