How do you handle context rot and belief revision in long-running LLM agents? (Building Corollary)

[gabe-santana]

🏷️ Discussion Type

Product Feedback

Body

Hey everyone!

So I've been building long running AI agents lately and kept bumping into context rot. Basically when past inputs or contradicted facts get stuck in the chat history, the LLM starts hallucinating because of stale context.

To try and fix this, I started working on Corollary. It's an open source python agent runtime based on Jon Doyle's 1979 paper on Truth Maintenance Systems (TMS).

Instead of appending messages line by line in a log, it treats agent state as a dependency graph of beliefs. So when a base fact changes, downstream data gets automatically invalidated and re-derived.

Heres a quick demo GIF:

Its written in Python.

Repo is here: https://github.com/gabe-santana/corollary

Would love ❤️ to hear how you guys manage state in complex agent workflows! Especially curious how people handle natural language contradictions when mapping graph dependencies.

Also, if anyone wants to jump in and contribute, I will tag a couple good first issue tickets on github for simple stuff.

Let me know what you think or if you have any feedback!

Guidelines

• I have read and understood this category's guidelines before making this post.

[github-actions[bot]]

💬 Your Product Feedback Has Been Submitted 🎉

Thank you for taking the time to share your insights with us! Your feedback is invaluable as we build a better GitHub experience for all our users.

Here's what you can expect moving forward ⏩

• Your input will be carefully reviewed and cataloged by members of our product teams. 
  • Due to the high volume of submissions, we may not always be able to provide individual responses.
  • Rest assured, your feedback will help chart our course for product improvements.
• Other users may engage with your post, sharing their own perspectives or experiences. 
• GitHub staff may reach out for further clarification or insight. 
  • We may 'Answer' your discussion if there is a current solution, workaround, or roadmap/changelog post related to the feedback.

Where to look to see what's shipping 👀

• Read the Changelog for real-time updates on the latest GitHub features, enhancements, and calls for feedback.
• Explore our Product Roadmap, which details upcoming major releases and initiatives.

What you can do in the meantime 💻

• Upvote and comment on other user feedback Discussions that resonate with you.
• Add more information at any point! Useful details include: use cases, relevant labels, desired outcomes, and any accompanying screenshots.

As a member of the GitHub community, your participation is essential. While we can't promise that every suggestion will be implemented, we want to emphasize that your feedback is instrumental in guiding our decisions and priorities.

Thank you once again for your contribution to making GitHub even better! We're grateful for your ongoing support and collaboration in shaping the future of our platform. ⭐

[varsh23-p]

Managing State & Natural Language Contradictions with TMS in LLM Agents

Hi @gabe-santana! Bringing Jon Doyle’s 1979 Truth Maintenance System (TMS / JTMS) into modern LLM agent runtimes is a brilliant approach to solving context rot and the lost-in-the-middle hallucination problem.

Here are a few practical architectural strategies for handling natural language contradiction detection and belief revision when mapping graph dependencies:

---

1. Converting Unstructured Text to Atomic Assertions (SPO Triples)

Mapping raw natural language directly into a TMS dependency graph is tricky due to ambiguity. A reliable pattern is decomposing conversation turns into Atomic Assertions / Semantic Triples with timestamps before feeding them to the belief graph:

User Turn: "Actually, let's switch from PostgreSQL to MongoDB for this project."
↓
Extraction:
- Retract: (project_db, type, PostgreSQL) [timestamp: T1]
- Assert:  (project_db, type, MongoDB)    [timestamp: T2]
- Dependency: [backend_schema, ORM_choice] depends on (project_db)

2. Handling Contradictions: Small NLI Models vs. Structured Predicates

When determining if Fact B contradicts Fact A:
Lightweight Natural Language Inference (NLI): Running a fast cross-encoder / NLI model (or small SLM) to classify [Entailment | Neutral | Contradiction] between new assertions and existing belief nodes.
Predicate Mutation Keys: Grouping beliefs under specific keys/namespaces (e.g., user.preferences.database_choice). Any update to the same key automatically triggers a contradiction check and marks the predecessor node as OUT.

3. Non-Monotonic Invalidation & Selective Context Hydration

The biggest advantage of Doyle's JTMS (Justification-based TMS) for LLMs is Selective Context Hydration:
Nodes hold justifications based on IN (believed) and OUT (disbelieved) sets.
When a root premise is retracted, downstream derivations are flipped to OUT without wiping their history.
When compiling the prompt context for the LLM turn, you only serialize the active IN subgraph, completely purging stale or contradicted history from the prompt window.

4. Useful Hybrid Frameworks for Comparison

If you're looking for reference architectures in this space:
ATMS (Assumption-based TMS by de Kleer): Especially helpful if you want the agent to explore multiple hypothetical branches in planning.
MemGPT / Letta: Hierarchical memory management (working context vs. archival storage).
Zep / Graphiti: Temporally-aware knowledge graphs that update and invalidate edges over time.