Skip to content

CHANGELOG.md

Latest commit

 

History

History
180 lines (137 loc) · 8 KB

CHANGELOG.md

File metadata and controls

180 lines (137 loc) · 8 KB

Changelog

All notable changes to AgenticMemory will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

v0.3.0 — V2: Grounding & Multi-Context Workspaces

Added

  • Grounding (anti-hallucination): memory_ground tool verifies claims have memory backing before an agent asserts them. Returns verified/ungrounded status with evidence.
  • Multi-context workspaces: Load and query across multiple .amem files simultaneously.
    • memory_workspace_create: Create a workspace for cross-memory queries.
    • memory_workspace_add: Load an .amem file into a workspace with role (primary/secondary/reference/archive).
    • memory_workspace_list: List all contexts in a workspace.
    • memory_workspace_query: Search across all loaded memories.
    • memory_workspace_compare: Find where a topic exists/doesn't across contexts.
    • memory_workspace_xref: Cross-reference with coverage summary.
  • 30 new V2 stress tests (grounding, workspace, integration).

Changed

  • SessionManager now includes WorkspaceManager for multi-context support.

[0.2.6] - 2026-02-23

Fixed

  • Added edge alias support for derived_from in MCP memory_add (mapped to caused_by) to improve cross-client compatibility.
  • Enabled memory_similar text-only fallback mode when embeddings are not available, removing a major usability gap.
  • Corrected memory_temporal MCP schema so range_a and range_b both strictly expose the typed RangeSpec contract.
  • Added regression tests for the new MCP alias and text-fallback behavior.

[0.2.4] - 2026-02-23

Fixed

  • MCP shutdown lifecycle hardening: stdio transport now exits immediately after a successful shutdown request.
  • Added protocol regression coverage to ensure shutdown acknowledgement is persisted and surfaced reliably.

[0.2.3] - 2026-02-23

Fixed

  • Publish parity fix: aligned agentic-memory and agentic-memory-mcp crate versions so crates.io verification uses the matching core API surface.

[0.2.2] - 2026-02-22

Fixed

  • Hardened stdio framing in MCP server paths for broader desktop/client compatibility.
  • Follow-up formatting and release hygiene updates in workspace crates.

Changed

  • Updated workspace documentation for orchestration behavior and install profiles.

[Unreleased] — v0.2.0 Remote Server Support

Planned

  • Remote HTTP/SSE transport (serve-http command)

    • --token flag for bearer authentication
    • --multi-tenant --data-dir for per-user brain files
    • /health endpoint for monitoring
    • --tls-cert / --tls-key for native HTTPS (optional)

  • New CLI commands

    • delete — remove a specific memory node
    • export — export brain to JSON
    • compact — defragment and optimize brain file

  • Infrastructure

    • Docker image (agenticrevolution/agentic-memory-mcp)
    • docker-compose with Caddy reverse proxy
    • Systemd service file
    • docs/remote-deployment.md

  • New error codes

    • UNAUTHORIZED (-32803), USER_NOT_FOUND (-32804), RATE_LIMITED (-32805)

Tracking: #1

[0.2.0] - 2026-02-19

Added

  • MCP Server (agentic-memory-mcp v0.1.0)

    • 12 MCP tools: memory_add, memory_query, memory_traverse, memory_correct, memory_resolve, memory_context, memory_similar, memory_causal, memory_temporal, memory_stats, session_start, session_end
    • 6 resources via amem:// URIs (node, session, type index, stats, recent, important)
    • 4 prompt templates: remember, reflect, correct, summarize
    • Stdio transport (default) + optional SSE transport (--features sse)
    • Session management with auto-save
    • Confidence validation at MCP layer (0.0-1.0 range enforcement)
    • Published to crates.io: cargo install agentic-memory-mcp

  • Monorepo restructure

    • Cargo workspace with crates/agentic-memory/ (core) and crates/agentic-memory-mcp/ (MCP server)
    • Integration bridge tests in tests/bridge/ (basic, multiagent, concurrent, stress)

  • Query Expansion: 9 new query types (queries 8-16)

    • BM25 text search with inverted index (fast path) and full-scan fallback (slow path)
    • Hybrid search combining BM25 + vector similarity via Reciprocal Rank Fusion (RRF)
    • Graph centrality: PageRank, degree centrality, and betweenness centrality (Brandes' algorithm)
    • Shortest path: bidirectional BFS (unweighted) and Dijkstra's algorithm (weighted)
    • Belief revision: counterfactual analysis with cascade propagation (read-only)
    • Reasoning gap detection: unjustified decisions, single-source inferences, low-confidence foundations, unstable knowledge, stale evidence
    • Analogical query: structural fingerprinting to find similar past reasoning patterns
    • Consolidation: deduplication, contradiction linking, inference promotion (with dry-run mode)
    • Drift detection: belief trajectory tracking with stability scoring

  • New index structures

    • TermIndex (tag 0x05): BM25 inverted index with posting lists
    • DocLengths (tag 0x06): dense array of token counts per node
    • Feature flags bitfield in header for forward/backward compatibility

  • 9 new CLI commands

    • amem text-search, amem hybrid-search, amem centrality, amem path
    • amem revise, amem gaps, amem analogy, amem consolidate, amem drift

  • Python SDK additions

    • 9 new Brain methods: search_text(), search(), centrality(), shortest_path(), revise(), gaps(), analogy(), consolidate(), drift()
    • New result dataclasses in agentic_memory/results.py

  • Backward compatibility

    • v0.1 files readable by v0.2 code (new queries use slow path)
    • v0.2 files skip unknown index tags gracefully for older readers
    • Feature flags in previously reserved header field

Test coverage

  • Rust core: 179 tests (83 new for v0.2 query methods)
  • MCP server: 119 tests (types, protocol, tools, resources, prompts, sessions, streaming, integration, edge cases)
  • Bridge integration: 16 tests (basic, multiagent, concurrent, stress)
  • Python SDK: 104 tests (20 new for v0.2 query wrappers)
  • Total across all suites: 575 tests

No new dependencies

All algorithms (PageRank, BFS, Dijkstra, BM25) implemented with std::collections only.

[0.1.0] - 2025-02-18

Added

  • Rust Core Engine

    • Binary graph format (.amem) with 6 cognitive event types and 7 edge types
    • LZ4-compressed content blocks
    • Memory-mapped I/O for zero-copy access
    • 128-dimensional feature vectors with cosine similarity search
    • Multi-level indexes (type, session, time, cluster)
    • CLI tool (amem) with create, add, link, info, traverse, query, mcp-serve commands
    • MCP (Model Context Protocol) server for IDE integration
    • 96 tests passing

  • Python SDK (pip install agentic-brain)

    • Brain class wrapping the Rust CLI with full API
    • MemoryAgent for LLM-powered agents with persistent memory
    • Provider integrations: Anthropic Claude, OpenAI GPT, Ollama
    • Automatic knowledge extraction from conversations
    • Context injection for memory-aware responses
    • 84 tests passing

  • Terminal Test Agent

    • Interactive agent with 6 validation protocols
    • Basic recall, decision recall, correction persistence, long-range memory, cross-topic inference, stress testing
    • 97 tests passing

  • Cross-Provider Validation

    • 21 tests validating memory portability across Claude, GPT, and Ollama
    • Binary format identity verification across providers

  • One-Command Installer

    • Auto-detection of 11 AI tools (Claude Code, Cursor, Windsurf, Continue, Ollama, LM Studio, etc.)
    • Automatic MCP configuration for supported tools
    • Backup and restore for all modified configs
    • 39 tests passing

  • Research Paper

    • 7-page publication-grade paper with 7 figures and 6 tables
    • Full benchmark data and methodology

  • Documentation

    • Quickstart guide, core concepts, API reference, file format specification
    • Integration guides for all supported tools and frameworks
    • FAQ and benchmark documentation