DREAM v3: Governing How Persistent Memory Influences AI Reasoning
For a long time, the primary memory problem in artificial intelligence appeared straightforward: how can an AI system store more context and retrieve it when needed?
As AI agents begin operating across longer periods, persistent memory becomes increasingly important. A useful agent should remember preferences, corrections, commitments, and relevant information across multiple sessions. However, persistence introduces a second problem that is less obvious and potentially more dangerous: A stored memory does not necessarily remain correct.
It may become outdated, be contradicted by new evidence, originate from an unreliable source, or accumulate distortions through repeated summarization. It may also be retrieved so frequently that the system begins treating repetition as evidence of truth.
This is the exact problem addressed by the third version of DREAM (Dynamic Retention Episodic Architecture for Memory).
The Central Thesis of DREAM v3: Retaining a memory does not automatically authorize that memory to influence current reasoning.
This distinction moves DREAM beyond adaptive memory retention and toward explicit governance of memory influence.
From Retention to Epistemic Stability
While DREAM v1 and v2 successfully established a scalable architecture to answer "How long should a memory remain available?", they left a major vulnerability unresolved. Access frequency measures use, but it does not measure truth, reliability, or temporal validity.
If the same signal governs both persistence and authority, the system may reinforce frequent errors while undervaluing rare but essential information. DREAM v3 formally separates these responsibilities into two independent policies:
- The Retention Policy: Answers "How long should this memory exist?" (Considering age, utility, redundancy, and privacy).
- The Influence Policy: Answers "Should this memory affect the current query, and with what authority?" (Considering confidence, contradiction pressure, temporal validity, and provenance integrity).
A memory may remain stored for historical, audit, or future re-evaluation purposes while having absolutely zero authority in the current context.
The Architecture of Governed Memory
To enforce this separation, DREAM v3 introduces several new architectural components designed to prevent unverified context injection.
1. Retrieval is NOT Authorization
Traditional systems rank records by similarity and inject the best results directly into a model’s context. In DREAM v3, retrieval merely completes the candidate-selection stage. A highly relevant memory may still be rejected if it is stale, contradicted, or derived from a broken compression chain.
2. Auditable Meta-Memory & Provenance Lineage
Controlling influence requires more than storing text. Every Episodic Unit now carries a meta-memory layer tracking its epistemic and operational state (e.g., Confidence, Contradiction score, Grounding status). Furthermore, Compression Lineage preserves the exact relationship between a semantic abstraction and the raw records it was derived from. A compressed conclusion can no longer silently replace its sources.
3. The Hygiene Gate
This is the single authorization point between retrieval and context construction. It combines a continuous Influence Score with explicit, non-compensatory rules to output strict actions:
- ALLOW: Injects the memory as authorized evidence.
- ALLOW_AS_UNCERTAIN: Injects the memory, but explicit uncertainty is appended.
- ARBITRATE: A material contradiction exists and must remain visible.
- LOW_INFLUENCE: The memory remains stored but is not normally injected.
- QUARANTINE: The memory is isolated due to epistemic risk.
- BLOCK: The memory is strictly prohibited from influencing context.
- CONSOLIDATE: Redundant records are queued for a lineage-preserving abstraction.
4. Preventing Recursive Self-Reinforcement
Persistent systems easily fall into a dangerous loop: a memory is retrieved -> it enters context -> the system interprets this exposure as utility -> the memory gains authority. DREAM v3 breaks this cycle with strict invariants:
- Retrieval does not count as confirmation.
- Context injection does not increase confidence.
- Repetition does not create truth.
- Model-generated restatements are not independent evidence.
Immutable Traces and Asynchronous Maintenance
Every consequential memory decision produces an Immutable Decision Trace. Serialized via canonical JSON and linked into a hash chain, it records exactly why a memory was authorized or prohibited from entering the context, referencing the specific policy version and features evaluated.
Meanwhile, an Asynchronous Maintenance plane works in the background to handle contradiction detection, evidence reconciliation, epistemic recalibration, and non-destructive pruning analysis, ensuring the database remains healthy without faking absolute certainty.
Empirical Evaluation: Controlled Results
DREAM v3 includes an executable Python reference core tested against a synthetic benchmark of 7,000 generated configuration runs. Comparing baseline systems (No Memory, Simple RAG, Permanent Memory, and a v2-style baseline), the complete DREAM v3 configuration achieved:
- 100% Grounding Fidelity
- 0% False Reinforcement (Blocked all forbidden injections)
- 0% Continuity Loss (Preserved all expected critical memories)
In our component ablation studies, removing the retention-influence separation caused the largest architectural failure, yielding a 34.3% false reinforcement rate. This provides controlled evidence that retention and influence must be independent architectural responsibilities.
Note on Limitations: These results demonstrate architectural behavior in a controlled synthetic environment. They do not claim that DREAM v3 universally determines truth, eliminates hallucinations natively, or alters the neural weights of an LLM. Influence here strictly means "controlled authorization of context."
Explore the Release
The v3 release moves the project from a conceptual proposal to an executable, auditable, and falsifiable research prototype. The repository includes the Reference Core (model-independent Python framework), synthetic benchmarks, ablation frameworks, and an interactive local playground for researchers to simulate intact or broken lineage, modify contradiction variables, and inspect hash-chained decision traces.
Project Links & Artifacts:
- GitHub Repository (DREAM v3 Source & Artifacts): MatheusPereiraSilva/dream-architecture
- Official Zenodo Record (DREAM v3): here
Acknowledgments: Special thanks to Patrick Morrison for conceptual discussions surrounding dynamic equilibrium, cognitive hygiene, and epistemic stability that helped shape this architecture.