flynn/docs/plans/phase3-pr1-adaptive-memory-compaction-checklist.md

Phase 3 PR #1 Checklist: Adaptive Memory Injection and Compaction Weighting

Created: 2026-02-12 Owner: Flynn core Status: ready to implement

Goal

Improve memory usefulness and compaction quality by injecting relevant memory context adaptively and preserving high-value turns during compaction.

Scope

In scope:

• adaptive memory relevance scoring
• configurable memory injection strategy
• compaction weighting for important turns (tool outcomes, corrections, preferences)
• tests and perf guardrails

Out of scope:

• new vector backend work
• UI controls for weighting

Files

New:

• src/memory/adaptive.ts
• src/memory/adaptive.test.ts
• src/context/weighting.ts
• src/context/weighting.test.ts

Modified:

• src/backends/native/orchestrator.ts
• src/context/compaction.ts
• src/memory/store.ts
• src/config/schema.ts

Implementation Steps

1. Add adaptive memory scorer:
   • keyword overlap with recent turns
   • recency weighting
   • token budget clipping
2. Add config flags:
   • memory.injection_strategy (all|recent|adaptive)
   • memory.max_injection_tokens
   • compaction.importance_threshold
3. Integrate adaptive injector in orchestrator memory injection path.
4. Add compaction message weighting and selection algorithm.
5. Ensure fallback to existing behavior on errors/timeouts.
6. Add tests for relevance selection and weighted compaction ordering.

Validation Commands

pnpm typecheck
pnpm test:run src/memory/adaptive.test.ts
pnpm test:run src/context/weighting.test.ts
pnpm test:run src/context/compaction.test.ts
pnpm test:run src/backends/native/orchestrator.test.ts
pnpm test:run
pnpm lint
pnpm build

Acceptance Criteria

• Adaptive mode injects fewer but more relevant memory snippets.
• Compaction preserves high-importance turns over low-value chatter.
• Token budgets are respected consistently.
• Existing behavior preserved when adaptive features disabled.
• No measurable latency regression beyond agreed budget.

Risks

• Relevance scoring false positives/negatives.
  • Mitigation: conservative scoring + extensive fixtures.
• Added latency in hot path.
  • Mitigation: bounded scoring time and fallback mode.

Commit Message

feat(memory): add adaptive injection and weighted compaction