neural-linking-memory-playbook.md

Neural Linking and Memory Playbook

A practical guide for using neural memory and neural-linking concepts in real agent systems.

1) What "neural linking" means for agents

In practice, neural linking is the combination of:

• retrieval linking (query -> relevant memory chunks)
• relational linking (entity/relation edges across memories)
• neural-symbolic linking (neural perception + explicit rules)

Goal: convert fragmented context into connected, actionable state.

2) Architecture pattern

Use a layered memory system:

1. Working memory (short horizon)

• Current objective, active plan, last tool outputs.

2. Episodic memory (session horizon)

• Compressed summaries and key decisions.

3. Semantic memory (long horizon)

• Stable facts, entities, relationships, procedures.

4. Retrieval layer

• Hybrid retrieval + rerank + link-aware packing.

3) Retrieval and linking pipeline

Recommended pipeline:

1. Normalize query (entities, time, domain terms)
2. Hybrid retrieve (lexical + vector)
3. Rerank by relevance and diversity
4. Build/expand local relation graph for top candidates
5. Pack context with citations and dependency order

Keep each retrieved chunk annotated with:

• source
• timestamp
• confidence
• relation targets (if available)

4) Neural + symbolic fusion pattern

Use neural models for:

• retrieval relevance
• semantic similarity
• uncertain inference

Use symbolic logic for:

• hard constraints
• safety/policy rules
• deterministic validation

Rule of thumb:

• probabilistic tasks -> neural
• compliance-critical tasks -> symbolic checks

5) Failure modes to watch

• Link drift: stale links between entities over time
• Retrieval pollution: unrelated chunks dominate context
• Over-compression: removes dependencies needed for reasoning
• Rule bypass: neural confidence overrides explicit constraints

Mitigations:

• link freshness checks
• duplicate suppression + rerank thresholds
• compression with protected fields
• mandatory symbolic validation on high-impact outputs

6) Evaluation checklist

• Multi-hop retrieval accuracy
• Entity linkage precision/recall
• Contradiction handling across linked memories
• Memory freshness under updates
• Constraint adherence after neural reasoning

7) Operator defaults

• Keep memory writes structured and sparse.
• Rebuild relation neighborhoods periodically.
• Prefer citations in final outputs.
• Treat long-context windows as fallback, not default design.

8) Core references

• Neural Turing Machines: https://arxiv.org/abs/1410.5401
• End-to-End Memory Networks: https://arxiv.org/abs/1503.08895
• Differentiable Neural Computer: https://arxiv.org/abs/1605.08582
• Transformer-XL: https://arxiv.org/abs/1901.02860
• Compressive Transformer: https://arxiv.org/abs/1911.05507
• RAG: https://arxiv.org/abs/2005.11401
• RETRO: https://arxiv.org/abs/2112.04426
• Neural Bellman-Ford Networks: https://arxiv.org/abs/2106.06935
• DeepProbLog: https://github.com/ML-KULeuven/deepproblog