1 · Recall Gate
Route before retrieval
Classify each question into memory types - exact fact, temporal update, persona, event graph, contradiction audit, abstention - and select ranked retrieval modes with confidence thresholds.
Memory metabolism · long-horizon recall
Astrocyte Memory
Not better RAG. Regulated cognition.
Long conversations are not a retrieval problem - they are a routing problem. Condensate’s Astrocyte layer decides which kind of memory a question needs before assembling context, verifies evidence after retrieval, and learns from answers - not just from co-occurring chunks.
The insight
Agents that run for weeks must handle multi-session memory, temporal updates, preference tracking, causal reasoning, summarisation, contradiction resolution, and abstention when evidence is missing.
Most systems fail because they retrieve the wrong memory type - a stale fact when the question needs the latest canonical state, or a similar chunk when the question needs an event chain.
Four regulatory layers
2 · Temporal hierarchy
Keep the path of change
Assertions carry validity windows and supersession links. Session summaries, persona policies, and event nodes form strata - not one flat summary blob.
3 · Evidence Verifier
Verify before answering
Post-retrieval checks: Is support direct? Is it temporally valid? Contradicted by later memory? Required source turns present? Abstain if not.
4 · Answer-aware feedback
Learn from correct paths
Reinforce retrieval paths that matched gold evidence - not merely paths that were retrieved together. Decay misleading connections after wrong answers.
Pipeline
Question ↓ Astrocyte Recall Gate ↓ Hybrid retrieval (canonical · temporal chain · persona · events · summaries) ↓ Rerank ↓ Astrocyte Evidence Verifier ↓ Answer (verified) or abstention ↺ Feedback loop (eval / POST /memory/feedback)
What Astrocyte addresses
| Memory challenge | Astrocyte response |
|---|---|
| Temporal (“what did X eventually decide?”) | Supersession chain + latest valid assertion |
| Multi-hop / causal | Event graph + graph spreading activation |
| Preference evolution | Persona stratum + temporal chain |
| Adversarial / trap answers | Abstention check + contradiction audit |
| Token budget | Mode-specific retrieval instead of full-transcript replay |
Positioning
Vector RAG optimises similarity. Astrocyte Memory optimises metabolism: raw experiences become assertions, assertions become temporal structures, structures become stable persona and event memory, and answer-time gates decide which pathway fires.
Condensate is built for agents that must remember accurately across sessions without shipping the entire transcript on every turn.
Technical reference: docs/architecture/astrocyte_memory.md in the Condensate repository.
API: POST /api/v1/memory/retrieve returns question_type, recall_plan, verification.