Finding

Paper: "Improving Coherence and Persistence in Agentic AI for System Optimization" (Engram, MIT)
arXiv: https://arxiv.org/abs/2603.21321 (March 2026)

Core Idea

Multi-run agent architecture solving context degradation over long horizons:

• Archive: persistent store of code snapshots + execution results across runs
• Research Digest: compact distillation of accumulated insights, injected into each fresh context
• Each new run starts with clean context but inherits all prior knowledge via the Digest
• Explicitly solves context window exhaustion + session recovery after failures

Applicability to Zeph (4/5)

Directly maps onto Zeph's existing architecture:

• Archive → Zeph's SQLite conversation history + semantic memory (Qdrant)
• Research Digest → Zeph's compaction summary + skill registry
• Fresh-context-with-digest → Zeph's context assembly with memory recall

Gap: Zeph has the components but not the explicit Digest distillation step. After hard compaction, the digest is implicitly the compaction summary. Making this explicit and persistent across sessions would enable true fault-tolerant long-horizon tasks.

Implementation Sketch

Add a table to SQLite: distilled key facts + outcomes updated at session end (similar to how compression guidelines are maintained). Inject the digest at context assembly time as a fixed top-of-context block, separate from per-turn memory recall. This is complementary to ACON compression guidelines (#2201) but session-scoped rather than compression-trigger-scoped.