Skip to content

MASTER-PLAN.md

Latest commit

 

History

History
69 lines (59 loc) · 4.46 KB

MASTER-PLAN.md

File metadata and controls

69 lines (59 loc) · 4.46 KB

Research Project: Solving the AI Context Window Problem

Master Plan & Progress Tracker

Created: 2026-01-31

Project Goal

Design an automatic system that preserves session context (tools, commands, decisions, plans) across AI coding assistant sessions, enabling seamless workflow continuity despite context window limitations. Produce a research paper documenting the full design process.

Research Methodology

Iterative convergence:

  1. Define the problem rigorously
  2. Survey existing approaches
  3. Brainstorm 5+ solutions
  4. Compare, rank, select top 2
  5. Deep-research top 2, generate new ideas
  6. Compare again — repeat until one solution is clearly superior
  7. Failure analysis on the winner
  8. Mitigate weaknesses
  9. Write research paper

Phase Tracker

Phase Status Output File Summary
1. Problem Definition ✅ COMPLETE phases/01-problem-definition.md 5 categories of lost info; 10 FRs, 8 NFRs; formal success criteria
2. Existing Solutions Survey ✅ COMPLETE phases/02-existing-solutions.md 15+ solutions surveyed; 9 design patterns; 5 gaps identified
3. Brainstorm Round 1 ✅ COMPLETE phases/03-brainstorm-r1.md 5 architectures: Journal, Palace, Git, Event Sourcery, Dual-Mind
4. Comparison Round 1 ✅ COMPLETE comparisons/04-comparison-r1.md Top 2: Event Sourcery (145/175) + Dual-Mind (143/175)
5. Deep Research + New Ideas ✅ COMPLETE phases/05-deep-research.md 3 hybrids: Cortex, Engram, Chronicle
6. Comparison Round 2 ✅ COMPLETE comparisons/06-comparison-r2.md Winner: Cortex (185/210), 14-pt margin over runner-up
7. (Not needed) SKIPPED — Cortex was clearly superior after Round 2
8. Failure Analysis ✅ COMPLETE phases/08-failure-analysis.md 19 failure modes; 2 critical, 6 high-risk
9. Weakness Mitigations ✅ COMPLETE phases/09-mitigations.md All risks reduced to ≤8/25; max reduction 87%
10. Paper Writing ✅ COMPLETE paper/cortex-research-paper.md Full research paper with 15 sections + appendices
11. External Evaluation ✅ COMPLETE evaluation/external-evaluation.md Independent stress-test: 8 holes, 10 missing items, priority-ranked
12. Evaluation Response ✅ COMPLETE evaluation/evaluation-response.md All P0/P1 items addressed; P2 items documented as limitations/future work

Key Decisions Log

  1. Event sourcing as foundation — Validated by industry consensus (BoundaryML, Akka, Graphite)
  2. No secondary LLM calls — Hard constraint from Phase 5 feasibility research
  3. Three-layer extraction — Structural + keyword + self-reporting covers >95% of events
  4. Progressive tiers — Critical mitigation for adoption risk (reduced 20/25 → 5/25)
  5. Immortal events for decisions — Decisions and rejections never decay
  6. .claude/rules/ for injection — Additive, never modifies user's CLAUDE.md
  7. SQLite + FTS5 + sqlite-vec — Single-file hybrid search, zero external dependencies
  8. [MEMORY:] tags for self-reporting — Most accurate extraction layer, trivially parseable
  9. Hook API validated — All 4 hooks (Stop, PreCompact, SessionStart, UserPromptSubmit) confirmed against official Claude Code docs with required payloads (Phase 12)
  10. Immortal events with growth management — Tiered briefing inclusion (active/aging/archived) resolves unbounded growth conflict (Phase 12)
  11. Bounded reality anchoring — Deterministic checks against structured data (git, config files, filesystem), not open-ended NLP matching (Phase 12)
  12. Evaluation-first implementation — Baseline collection before Cortex, instrumented metrics, A/B comparison after Tier 0 (Phase 12)
  13. Decay parameters are tunable — All thresholds configurable and calibrated from real sessions, not hardcoded magic numbers (Phase 12)

Cross-Session Context Index

Last Updated: 2026-01-31

Current Phase: RESEARCH COMPLETE — READY FOR IMPLEMENTATION

Key Artifacts Created:

  • 6 phase documents in phases/
  • 2 comparison documents in comparisons/
  • 1 comprehensive research paper in paper/ (updated with evaluation response)
  • 1 external evaluation in evaluation/
  • 1 evaluation response in evaluation/

Critical Decisions Made: See Key Decisions Log above (13 decisions total)

Winner: Cortex — Event-Sourced Memory with Projected Briefings

Next Action: Implementation of Tier 0 prototype (start with baseline collection per §11.4)