Forge Decision Context Turns Corrections into Persistent Agent Rules
Summary:Forge Decision Context (forge_dc) addresses a central limitation of LLM agents: corrections rarely persist across separate sessions. The system extracts imperative RULE directives from human feedback and self-assessed tool-call failures, then injects only the rules relevant to the current decision point. On DC-Bench, a new benchmark containing 71 scenarios across nine behavioral dimensions, forge_dc reports 95% compliance, compared with 55% for naive retrieval, while injecting approximately ten
The Problem: Agents Remember Facts but Not Behavior
LLM agents can persist user preferences, entity information, and conversation history across sessions, yet they often fail to retain procedural knowledge about how they should act. Users therefore have to repeat the same correction at the start of every new conversation. Conventional memory systems function more like searchable archives than executable behavioral policy.
Forge Decision Context (forge_dc) represents behavioral experience as explicit imperative rules. It extracts signals from human corrections and self-assessed tool-call failures, compiles them into scoped RULE directives, and injects only the rules relevant to the current decision point. The critical shift is the unit of memory: declarative facts become actionable constraints.
Representation, Activation, and Placement
The system can be understood through three engineering layers.
Representation. Behavioral information is compiled into imperative rules rather than conversational summaries or arbitrary vector chunks. Rules act as an execution interface for the agent and can express both prohibitions and required actions under specified conditions. This reduces the need for the model to rediscover operational intent from an ambiguous history.
Activation. The system determines which rules apply to the current task. Injecting every rule into every prompt raises token costs and may distract the model with irrelevant constraints. forge_dc activates rules at relevant decision points instead.
Placement. Rules are inserted where they can affect the relevant tool selection, parameter generation, or action step rather than being appended indiscriminately to the system prompt. This may improve adherence in long agent trajectories, although the benefit depends on context structure, positional encoding, and trajectory length.
At the systems level, forge_dc resembles decision-point-aware external state management. A rule store maintains persistent behavior, a router selects applicable directives, and the executor enforces them during tool use. It avoids model retraining, but introduces rule extraction, conflict resolution, version management, and runtime routing costs.
DC-Bench: 95% Compliance with Lower Context Overhead
The authors introduce DC-Bench, containing 71 scenarios across nine behavioral dimensions. The scenarios are derived from real executed agent traces and run in a deterministic mocked-tool environment to reduce variability from external services.
The paper reports 95% behavioral compliance for forge_dc, compared with 55% for naive retrieval. It also injects approximately ten times fewer tokens per query, suggesting that selective activation can improve rule use while reducing persistent-context overhead.
These results should still be interpreted within the evaluation boundary. A deterministic mocked environment improves reproducibility but does not reproduce production network delays, noisy tool responses, permission failures, or state drift. The abstract does not disclose model configurations, decoding parameters, conflict-handling policies, statistical significance, or cross-model generalization. The 95% figure is therefore a result reported on DC-Bench, not a universal guarantee for agent frameworks.
Engineering Trade-offs
Lower context cost does not imply low maintenance cost. A production deployment must address several issues.
First, rule extraction must be reliable. Human corrections may contain emotion, explanations, and conflicting demands. Tool failures can also reflect transient network failures or missing dependencies rather than incorrect behavior. Treating every failure as a durable rule risks polluting future sessions.
Second, rules need conflict and priority management. Records should include source, time, scope, confidence, priority, and expiry. Temporary corrections need short time-to-live values, while stale rules require explicit invalidation.
Third, activation introduces its own bottleneck. Reduced prompt size does not proportionally reduce total inference cost. Scaling the rule store may increase retrieval overhead or trigger irrelevant rules. Vector retrieval or model-based routing can add calls and index-maintenance costs.
Finally, safety requires traceability. Self-assessed failures are not reliable supervision by default. Persistent directives should support review, revocation, and deletion, especially in finance, operations, and enterprise-data agents.
Deployment Guidance
- Separate behavioral rules from factual memory. Use distinct schemas and retrieval policies for procedural constraints and user or entity facts.
- Attach scope and lifecycle metadata. Record source, applicable tools, trigger conditions, creation time, confidence, and expiration.
- Inject at high-risk decision points. Prioritize tool calls, permission requests, deletions, and irreversible actions over ordinary conversational turns.
- Preserve human confirmation. Confirm high-impact rules and assign short TTLs to one-off corrections.
- Regression-test with real traces. Measure compliance alongside false activation, rule coverage, token use, end-to-end latency, and recovery rate.
- Test cross-model transfer. If one model extracts rules and another executes them, measure semantic drift and formatting adherence separately.
The main contribution of forge_dc is not another generic memory store. It decomposes persistent behavioral learning into representation, activation, and placement, providing a lightweight path toward cross-session operational norms. Production readiness still depends on reliable extraction, conflict governance, and independent validation in realistic environments.
— END —Source: ArXiv AI (cs.AI) (2026-11-10)
Tags: #Forge Decision Context #Agent Memory #Behavioral Learning #Decision Context #DC-Bench
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments