Skip to content

FEATURE_ROADMAP.md

Latest commit

 

History

History
84 lines (67 loc) · 4.14 KB

FEATURE_ROADMAP.md

File metadata and controls

84 lines (67 loc) · 4.14 KB

Feature Roadmap

Evolving Engram from a memory store into a fully-fledged brain for AI agent persistence.

Completed

Fine-Grained Memory Editing

  • Edit by ID with update — direct targeting, no similarity guessing
  • Partial content edits: append, prepend, find+replace modes
  • Field-level updates: change topic, project, source, importance, expires_in_days without touching content
  • Re-embed on content change, skip for metadata-only changes

Knowledge Graph

  • edges table with (source_id, target_id, relation) — six relation types: relates_to, contradicts, supersedes, derived_from, part_of, summarized_by
  • connect / disconnect tools for creating and removing edges
  • graph tool to view a memory's neighborhood at configurable depth
  • recall with depth: 1-3 follows edges via BFS to surface connected knowledge
  • Hierarchical memories via parent_id (auto-creates part_of edges)
  • Edges auto-cleanup on forget/merge

Reinforcement / Importance Scoring

  • accessCount and importance fields on Memory model
  • Scoring: cosine similarity blended with frequency (log-scaled, 15% boost), importance (1-5, 20% boost), recency (exponential decay, 10% boost)
  • Recall ranking uses reinforcement signals to fine-tune ordering among close matches

Conflict Detection

  • remember checks for near-duplicates using both cosine distance (< 0.12 same-project, < 0.05 cross-scope) AND Jaccard term overlap (40%+ shared terms)
  • Blocks storage with a warning showing the existing memory; force: true to override

Hybrid Search (Semantic + Full-Text)

  • FTS5 full-text index on content field
  • Recall combines FTS5 (any matching term) with vector cosine similarity
  • Falls back to FTS5-only if embedding model is unavailable

Episodic Memory

  • Episode model with title, summary, project, timestamps
  • begin_episode / end_episode / recall_episode / list_episodes tools
  • Auto-episodes: first remember creates one; >30 minute gap or project switch starts a new one

Task Continuity

  • Checkpoint model with plan, progress, context, status
  • checkpoint / resume / list_tasks tools
  • Tasks persist across conversations with active/paused/completed status

Temporal Queries

  • timeline tool with chronological view grouped by day/week/month
  • since / before temporal filters on recall
  • Calendar-style grouping with project and topic filters

Clustering & Consolidation

  • find_clusters uses greedy cosine-distance + Jaccard term overlap clustering
  • consolidate creates a summary memory, deprioritizes originals (importance -> 0), links with summarized_by edges

Memory Visualizer

  • Interactive force-directed graph (SwiftUI Canvas, 60fps)
  • Project-colored nodes with inter-project repulsion for visual separation
  • FTS5-backed search with prefix matching, highlight mode
  • Edge type filtering with color-coded relations
  • Time slider with play button to animate graph growth
  • Semantic cluster hulls, minimap, detail panel, keyboard shortcuts
  • PNG export

Not Yet Implemented

Confidence Levels and Provenance Tracking

  • confidence field (0.0-1.0) that increases with confirmation across sessions, decreases with conflicting evidence
  • Provenance chain tracking which conversations/files/events created or reinforced a memory
  • Low-confidence memories flagged for verification

Proactive Surfacing / Triggers

  • Register trigger patterns that auto-surface memories when context matches
  • Safety triggers ("never force-push to main"), convention triggers ("this project uses tabs")
  • add_trigger / remove_trigger tools

Multi-Type Memory

  • Typed memories: fact, code_pattern, error_resolution, decision, preference, procedure
  • Structured fields per type (e.g., error_resolution has symptom, root_cause, fix, prevention)
  • Type-aware retrieval and display formatting

Memory Validation / Truth Maintenance

  • Verify stored memories against current codebase state (file paths, function signatures, dependency versions)
  • Auto-validation on project recall, spot-checking verifiable memories
  • Stale memories flagged or auto-expired
  • Git integration for detecting significant file changes