Context Regeneration & Perfect Recall

Status: stable | Score: 100% | Last Tested: 2026-07-18

Overview

Context Regeneration and Perfect Recall implements an automated context assembly and verification pipeline. By processing real message sequences, it evaluates and optimizes Recall@K and MRR (Mean Reciprocal Rank), rebuilding working memories to maximize retrieval accuracy under varying token limits.

Architecture & Design

The context pipeline relies on a dedicated WorkingMemory which acts as a real-time, volatile cache of active agent context. If a query requires deeper details, the “turn pack” virtualizes the conversational history, performing targeted episodic retrieval to pull relevant past turns and stitch them together using unified token estimations.

Implementation Paths

• src/context/ (WorkingMemory structures and turn packaging)
• src/memory/ (episodic retrieval, candidate scoring, and context builders)

Sub-features

• ctx-recall-harness: Measures Recall@K on production benchmarks and flags quality degradation.
• ctx-turn-pack: Packages multi-turn conversation histories into tiered context blocks.
• ctx-episodic-real: Extractive episodic summarizer that runs automatically on message stores.
• ctx-working-wire: Direct integration of the volatile WorkingMemory with active agent completion steps.

Test References

• Working memory updates and episodic retrieval precision tests.
• Recall harness correctness and score tracking unit tests.

Known Issues & Notes

• Bounded to 100 most recent documents (the hot set) during retrieval steps to prevent full database dumps.
• Context compression algorithms are fully functional, with subsequent polish added incrementally.

Functional Context Regeneration Example

Regenerate agent context using Turn Packs to maintain peak semantic recall:

curl -X POST "http://localhost:8006/v1/context/regenerate" \
  -H "X-Xavier-Token: $XAVIER_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "sess_91823",
    "compression_mode": "extractive",
    "target_token_budget": 2048
  }'