Add time-based decay, anti-patterns, and agent attribution

[acoyfellow]

Summary

This PR adds three major features to deja-local and deja-edge memory systems:

1. Time-based confidence decay — memories automatically lose confidence over time using an exponential half-life formula, with decay computed at recall time (read-side only)
2. Anti-pattern tracking — memories rejected enough times auto-invert into warnings that actively surface during recall
3. Agent attribution — optional source parameter to track which agent stored a memory

Key Changes

Time-based Confidence Decay

• Added last_recalled_at column to track when memories were last used
• Decay formula: decayedConfidence = storedConfidence × 0.5^(daysSince / 90)
• A memory untouched for 90 days has its effective confidence halved
• Recalling a memory resets its decay clock — actively used knowledge stays fresh
• Decay is computed at recall time; stored confidence values are never mutated
• No background jobs or cron needed — pure read-side computation

Anti-Pattern Tracking

• Added type column ('memory' or 'anti-pattern') to track memory classification
• When reject() drops confidence below 0.15 threshold, memory auto-inverts:
  • Type changes from 'memory' to 'anti-pattern'
  • Confidence resets to 0.5 (useful warning, not failed memory)
  • Text is prefixed with "KNOWN PITFALL: "
• Anti-patterns appear in recall results normally, actively warning agents
• Already-inverted anti-patterns don't double-invert on further rejections
• Anti-patterns can be confirm()ed to boost their confidence

Agent Attribution

• Added source column to store which agent created a memory
• remember() now accepts optional { source?: string } parameter
• Source is returned in Memory objects and list results
• Backward compatible — source is undefined when not provided

Schema Migrations

• Added migration logic to handle DBs created before these features
• Gracefully adds missing columns: last_recalled_at, source, type
• Existing memories default to type: 'memory' and source: undefined

Documentation Updates

• Updated README and pattern guides with decay mechanics and anti-pattern examples
• Added hosted service decay example in memory-hygiene.mdx
• Documented agent attribution usage

Implementation Details

• deja-local: Uses Bun SQLite with prepared statements; decay computed during recall
• deja-edge: Uses Cloudflare Durable Objects SQLite; same decay logic applied
• Hosted service: Added decay ranking in injectMemories() to sort learnings by decayed confidence
• All changes maintain backward compatibility with existing databases
• Comprehensive test coverage added for all three features in both packages

https://claude.ai/code/session_01MbQ14BcEn9uFntgrNidpVA

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e25870fb5d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".