Skip to content

Changelog

All notable changes to engram. Follows Keep a Changelog conventions.

[v0.5.25] — Current

Latest release. See GitHub releases for the full release history.

[v0.4.3] — 2026-03-01

Interactive Setup Wizard

Added

  • engram setup command — interactive wizard to auto-detect and configure AI agents/IDEs
  • Auto-detects: Claude Code, OpenClaw, Cursor, Windsurf, Cline, Aider, Zed, Void, Antigravity
  • --dry-run flag previews changes without writing
  • --non-interactive flag for CI/headless environments

  • --status flag shows current connection status

  • Connector architecture (src/engram/setup/connectors/)

  • AgentConnector ABC with detect(), configure(), verify() interface
  • McpJsonConnector mixin for MCP JSON config merge

  • Connector registry with auto-discovery, sorted by tier (1-3)

  • MCP path resolution fix — resolves engram-mcp to absolute path when installed in a venv

  • Federation provider stubs for Mem0, Cognee, Zep detection

[v0.4.2] — 2026-02-28

Configuration Management + Live Settings Editor

Added

  • Configurable everything — all hardcoded tuning parameters now YAML-configurable:
  • ExtractionConfig: llm_model, temperature, max_retries, chunk_size
  • RecallConfig: search_limit, entity_search_limit, graph_depth, scoring params

  • SchedulerConfig: consolidate/cleanup/decay intervals, task timeout

  • WebUI Settings Editor (new /Settings tab in web UI)

  • GET/PUT /api/v1/config endpoints for read/write config.yaml
  • Grouped config sections: LLM, Embedding, Recall, Extraction, Scheduler, Health, Cache
  • Live preview with type validation and restart-required badges

[v0.4.1] — 2026-02-25

Memory Classification + Actionable Memories

Added

  • capture/memory_classifier.py — heuristic classifier (regex patterns, no LLM cost)
  • Classifies: todo, decision, preference, error, workflow, lesson, fact
  • Supports English and Vietnamese patterns
  • Integrated into do_ingest() and do_ingest_messages() pipelines

[v0.4.0] — 2026-02-20

Entity-First Ingestion

Changed

  • Switched to /api/v1/ingest endpoint for entity-first ingestion
  • EntityExtractor now gates storage — only messages with extracted entities are stored
  • Dual storage: episodic (Qdrant) + semantic graph updated atomically on each ingest

[v0.3.2] — 2026-02-15

Brain Features + Intelligence Layer

Added

  • Temporal resolution — 28 Vietnamese+English date patterns
  • Pronoun resolution — graph-context + LLM fallback
  • Fusion formatter — [type]-grouped recall output
  • Memory consolidation — Jaccard clustering + LLM summarization
  • Meeting ledger — structured meeting records
  • Feedback loop — confidence scoring (+0.15/-0.2), auto-delete on 3x negative
  • Graph visualization — interactive vis-network UI with dark theme
  • Data constitution — 3-law LLM governance, SHA-256 tamper detection

[v0.3.1] — 2026-02-10

Recall Pipeline

Added

  • Full recall pipeline: query decision → temporal resolution → pronoun resolution → parallel search → dedup → fusion
  • Auto-memory detection — save-worthy message detection without explicit engram remember
  • Poisoning guard — blocks prompt injection from being stored
  • Retrieval audit — JSONL log for all recall operations
  • Auto-consolidate trigger after N messages

[v0.2.0] — 2026-02-01

Federation Layer

Added

  • Provider adapters: REST, File, PostgreSQL, MCP
  • Auto-discovery: local port scan + remote hosts + MCP config files
  • Smart query router: internal vs domain classification
  • Circuit breaker: auto-disable on consecutive errors
  • Plugin system: third-party adapters via entry_points

[v0.1.0] — 2026-01-15

Initial Release

Added

  • Dual-memory system: EpisodicStore (Qdrant) + SemanticGraph (NetworkX + SQLite/PG)
  • Reasoning Engine (Gemini via litellm)
  • CLI (Typer), MCP server (FastMCP), HTTP API (FastAPI), WebSocket API
  • Multi-tenancy, JWT auth, Redis caching, rate limiting
  • Audit logging, OpenTelemetry, Docker deployment
  • Backup/restore, health checks, benchmark suite