Source-linked AI summary
AkasicMEM: Governed Enterprise Memory for Agents
Jeongmin Bae, Yongjae Kim, Kyoung Hur, Donghyoung Han, Min-Soo Kim
TL;DR
Enterprise agent memory must integrate enterprise sources with governed shared memory, while preserving source restrictions as information is repeatedly derived and reused under changing principals and policies. AkasicMEM implements this authorization continuity with lineage, policy composition during formation, and policy re-evaluation at retrieval, as a realization of Governed Enterprise Memory.
Problem
The paper addresses the lack of a combined lifecycle design for source–memory integration, memory governance, and authorization continuity, despite their interaction when enterprise-source information persists in memory.
Source: §1 · pp. 1–2Method
Results
AkasicMEM realizes authorization continuity by carrying source restrictions through memory derivation and reuse under changing policies and principals.
Source: §6 · p. 8Takeaways & Limitations
Governed Enterprise Memory treats authorization as a lifecycle-wide dependency structure managed across source and memory derivation and reuse.
Source: §6 · p. 8The automated declassification test is necessary but not sufficient to establish that a derived memory no longer depends on restricted information.
Source: §5.2 · p. 8
Abstract
from arXiv · showhide
Agent memory enables enterprise agents to retain knowledge acquired during work and reuse it across tasks and agents, turning execution experience into persistent organizational knowledge. Realizing this potential requires both source--memory integration, through which enterprise sources and accumulated memory can be utilized together, and memory governance, through which shared memory remains subject to organizational policies throughout its lifecycle. These requirements interact when information from enterprise sources persists in memory. As this information is repeatedly derived and reused under changing principals and policies, source restrictions may be bypassed, resulting in information leakage. Preventing such leakage requires authorization continuity, under which source restrictions remain effective throughout source-to-memory and memory-to-memory derivation and reuse. Existing approaches address these concerns individually, but do not treat source--memory integration, memory governance, and authorization continuity as combined core design targets across the memory lifecycle. We define Governed Enterprise Memory as agent memory designed around this combined scope and present AkasicMEM as its realization. AkasicMEM realizes authorization continuity through transitive lineage, policy composition during memory formation, and policy re-evaluation during retrieval. It is built on GraphAI's AkasicDB, a unified vector--graph--relational database whose storage and execution substrate enables the underlying operations of these mechanisms to be jointly optimized and executed.
1 Introduction
Enterprise agent memory can preserve execution knowledge for reuse, but realizing that value requires integrating enterprise sources with governed shared memory. Because source-derived information can persist through repeated derivation while principals and policies change, the paper defines authorization continuity and presents AkasicMEM to maintain it across the memory lifecycle.
1 Introduction
- Enterprise agent memory complements explicitly specified skills by accumulating facts, decision context, and execution outcomes learned during work for subsequent tasks.Reported deployments and multi-agent work illustrate reuse of organization-specific knowledge and collaboration lessons [8] [34].Source: §1 · p. 1
- Source–memory integration combines enterprise-source information with accumulated memory while preserving their distinction, so task context can draw on both current business states and prior task knowledge.Retrieval over enterprise sources, commonly implemented through RAG [18], must be coordinated with memory formation and retrieval.Source: §1 · p. 1
- Source restrictions must remain effective as memory is repeatedly derived and reused under changing policies and principals, or memory scope alone can permit information leaks.The paper calls this requirement authorization continuity.Source: §1 · p. 1
- AkasicMEM treats source–memory integration, memory governance, and authorization continuity as combined lifecycle design targets, using lineage, formation-time policy composition, and retrieval-time policy re-evaluation.It records declassification and retrieval decisions for audit; its technical foundation is AkasicDB, which supports vector, graph, and relational processing on one execution substrate [16].Source: §1 · p. 2
2 Background and Related Work
The paper frames agent memory as a lifecycle-managed store of persistent information, distinct from read-oriented retrieval and context techniques. Existing systems target different reuse boundaries; the cited approaches do not combine source integration, governance, and authorization continuity across the lifecycle.
2.1 Agent Memory and Its Lifecycle
- Agent memory comprises persistent information used in later decisions, organized as semantic, episodic, or procedural memory objects managed throughout their lifecycle.The paper distinguishes this long-term memory from working memory for the current task [13] [27].Source: §2.1 · p. 2
- The paper treats forgetting as a distinct lifecycle dynamic because retention expiry, contract termination, and source revocation each trigger removal with its own obligations.Formation, retrieval, evolution, and forgetting are the four dynamics adopted in this account [13] [32] [15].Source: §2.1 · p. 2
- Unlike conventional RAG, agent memory decides what becomes persistent, reuses it across sessions and tasks, and updates or removes it as information changes.Long-context processing and context management can instead serve working-memory functions [3] [35] [18] [13].Source: §2.1 · p. 2
2.2 Existing Agent Memory Systems
- The comparison groups representative systems by their primary reuse boundary and rates them against the three enterprise-memory requirements, with ratings reflecting stated design scope rather than every product feature.The table’s categories are core, partial, and outside stated scope.Source: §2.2 · pp. 2–3, Table 1 · p. 3
- Governed shared-memory systems track provenance or evaluate current access conditions, but the paper identifies remaining gaps in common source–memory management and full authorization continuity.The discussed gaps include preserving derivation dependencies, composing inherited restrictions, and re-evaluating ancestor policies as permissions change.Source: §2.2 · p. 3
3 Governed Enterprise Memory
Governed Enterprise Memory combines source–memory integration with memory governance, focusing here on keeping source authorization effective through repeated derivation and changing principals or policies. AkasicMEM’s requirements formalize that continuity through lineage, composed restrictions, and current-policy checks.
3.1 Authorization of Source-Derived Memory
- A shared-agent memory scope can let Bob retrieve information derived from Alice-authorized source data even after the source denies Bob access.The example distinguishes permission to reuse memory within the agent’s scope from authorization to access information inherited from its source.Source: §3.1 · p. 4
3 Governed Enterprise Memory
- Authorization continuity requires inherited restrictions to keep governing derived memory across repeated derivation and changes in principals and managed policy state.The paper positions this as the focus of Governed Enterprise Memory alongside broader governance concerns such as validity, retention, deletion, and auditability.Source: §3 · p. 3, §3.2 · p. 4
3.2 Authorization Continuity
- Transitive lineage records source and memory contributions at formation and preserves derivation links so ancestors remain computable through further derivation.This retains provenance and an auditable account of how derived memory was formed.Source: §3.2 · p. 4, Figure 1 · p. 4
- A new memory combines its intended reuse scope with restrictions inherited from its contributing sources and memories, then retrieval re-evaluates those restrictions against current principals and policy state.Re-evaluation must account for both revoked and newly granted authorization; declassification is a recorded, sanctioned exception to inherited restrictions.Source: §3.2 · p. 4
4 AkasicMEM
AkasicMEM is a governed memory layer that manages enterprise sources and memory across a shared lifecycle. It maintains authorization continuity through lineage and policy composition at formation, policy re-evaluation at retrieval, and unified execution on AkasicDB.
4.1 System Overview
- AkasicMEM manages governed source and memory objects linked by lineage, while lifecycle operations form, retrieve, evolve, and forget memory in shared storage.Source: §4.1 · pp. 4–5
4.2 Governed Memory Formation
- Store records direct source and memory lineage and composes policy from scope plus every input restriction; explicit declassification can exclude an ancestor while retaining lineage for audit.Source: §4.2 · p. 6
4.3 Governed Memory Retrieval
- Policy changes flag potentially affected downstream memories; when a flagged candidate is retrieved, AkasicMEM recomputes authorization from current restrictions and serves it only if the requesting principal qualifies.Source: §4.3 · p. 6
4.4 Unified Execution on AkasicDB
- AkasicDB jointly plans vector selection, relational filtering, lineage traversal, and policy evaluation under a common transaction, avoiding cross-system coordination without removing the cost of those operations.Source: §4.4 · pp. 6–7
5 Discussion and Future Work
The discussion treats lineage as the structure for extending governance beyond authorization continuity, while identifying lifecycle boundaries and evaluation needs.
5.1 Governance Beyond Authorization Continuity
- Source corrections can leave derived memory authorized but false, so lineage should flag descendants for revision or invalidation and preserve revisions as new versions.Resolving content changes requires re-verification against evidence or a person; policy changes, by contrast, compose mechanically.Source: §5.1 · p. 7
- Retention and deletion obligations must propagate through lineage, but multi-source derivations require judgment about continued dependence and semantic probing to verify forgetting.Derived memory should inherit the most restrictive source retention period; deliberate declassification can preserve a useful lesson after its source expires.Source: §5.1 · p. 7
- Reproducing past access decisions requires versioned authorization graphs and source policies, while recorded decisions can identify recipients of disclosures after revocation.Audit records themselves need authorization because they reveal the existence and reach of memory.Source: §5.1 · p. 7
5.2 Toward Useful Governed Memory
- Selection remains outside scope; governed formation favors an observer assigning authorization from lineage, while consolidation across sources accumulates their restrictions.The proposed practical pattern filters early, stores observations near evidence, and consolidates offline; integration of consolidation into governed formation remains ongoing work.Source: §5.2 · p. 7
- Current retrieval selects the most query-similar memory a principal may see under current policy, but does not yet use demonstrated usefulness to rank or forget memories.Retrieval signals could identify repeatedly relevant memories blocked by governance for declassification review.Source: §5.2 · p. 8
- Controlled declassification can widen a derived memory’s audience when it no longer depends on restricted material, but automated inference is only a screen for human approval.Sentence-level independence may still disclose restricted information in aggregate.Source: §5.2 · p. 8
- Outcome feedback is not yet implemented, and recording whether memory helped a principal must itself remain content-free and scoped because it can disclose cross-engagement use.Source: §5.2 · p. 8
5.3 Evaluating Governed Enterprise Memory
- Evaluation must vary sources, principals, derivation, and governance events, measuring leakage, over-restriction, relevance, utility, freshness, trust, and overhead against baselines.Semantic leakage requires judged comparison or adversarial probing, and results must expose trade-offs among leakage, over-restriction, and cost rather than a single score.Source: §5.3 · p. 8
6 Conclusion
The conclusion defines authorization continuity for memory derived from enterprise sources and presents AkasicMEM as its realization. It frames lifecycle-wide lineage and the tension between accumulated restrictions and shared knowledge as central to Governed Enterprise Memory.
6 Conclusion
- Authorization continuity keeps source restrictions effective as memory is repeatedly derived and reused under changing policies and principals.Source: §6 · p. 8
- AkasicMEM records lineage at formation, composes policies from sources, and re-evaluates them at retrieval over integrated source–memory storage on AkasicDB.Source: §6 · p. 8
- Governed Enterprise Memory is framed as a lifecycle-wide dependency structure, with controlled declassification addressing the tension between accumulated restrictions and shared-knowledge value.Source: §6 · p. 8