fix(reflection): require explicit staleness before overwriting memory facts

[Tusharv23]

Summary

Addresses #4029.

The sleeptime reflection prompt currently instructs the subagent to "resolve contradictions in favor of the latest evidence" unconditionally. This newest-wins rule causes destructive memory overwrites: a single unverified or anecdotal observation can replace a confirmed durable fact, including safety-critical ones (e.g., a confirmed severe allergy softened to "mild" after one anecdote about eating shellfish without reaction).

This PR replaces the inline conflict clause with a structured Durability and Resolutions policy: update only when newer evidence explicitly establishes staleness; otherwise record the conflicting observation as a separate fact and leave the original unchanged; never weaken confirmed safety-critical facts on anecdote.

Prompt-only change (src/agent/subagents/builtin/reflection.md); no code touched.

Tradeoff: memory grows slightly faster (conflicts are annotated instead of overwritten), in exchange for not destroying confirmed facts on weak evidence.

Verification

• Measured on a 120-run eval matrix: 4 conflict cases (true update / compatible split / inferential overwrite / confidence preservation) x 10 Bedrock models (frontier + small) x 3 repeats, two-stage harness (verbatim reflection prompt, then neutral extractor, then mechanical scoring with manual adjudication of edge cases):
  • Baseline reflection.md: 90/120 correct conflict handling
  • This PR's prompt: 110/120
  • True-update case stays 100%, so legitimate updates still go through
  • Remaining failures concentrate in the smallest models (llama3.1-8b, nova-micro)
• The exact committed text was additionally sanity-checked (N=12: 3 models x 4 cases) and behaves identically to the measured version.
• Full harness, raw run evidence, and scoring code: https://github.com/Tusharv23/letta-conflict-eval
• bun test on this branch: 6872 pass / 46 fail, identical counts to clean main on the same machine (pre-existing environment-dependent failures: telemetry auth, PTY, stream-json integration). No new failures introduced.

AI Disclosure

• This pull request was written entirely by a human
• This pull request was written with AI assistance and reviewed and edited by a human
• I have read the AI Policy and agree to its terms

AI Tool(s) Used

GitHub Copilot (Claude Fable 5) - used to help draft the prompt wording, plan the eval harness. All runs, changes and eval results done by me.

Human Verification

I have reviewed and understand every change in this pull request and take responsibility for its correctness.

[Tusharv23]

Hey @kl2806, checking in on this PR. Since submitting I've re-run the sanity set against the final wording (11/12, results in the PR body) and added confidence intervals to the underlying eval. Happy to make any changes if the wording or scope needs adjusting, is there anything blocking a review?