agent-recall-core

Implements a hierarchical memory palace architecture that organizes agent interactions and knowledge into spatially-indexed semantic rooms. Uses a graph-based storage model where each 'room' represents a conceptual domain, with memories encoded as nodes connected by semantic relationships. The system maps abstract information to spatial locations, enabling agents to retrieve contextually relevant memories through spatial navigation rather than keyword search.

Exposes memory palace functionality as MCP (Model Context Protocol) tools, allowing Claude and other MCP-compatible agents to interact with the memory system through standardized tool calling. Implements MCP resource handlers for memory read/write operations, with schema-based function definitions for memory operations like store, retrieve, navigate, and update. Enables seamless integration with Claude's native tool-use capabilities without custom client code.

Handles conflicts when multiple agents or processes write to the same memory simultaneously, using configurable merge strategies (last-write-wins, semantic merging, manual conflict resolution). Detects conflicting updates to memory nodes and applies merge logic to reconcile differences while preserving important information. Supports both automatic merging (for non-conflicting updates) and manual conflict resolution (for semantic conflicts).

Implements multi-criteria memory retrieval that ranks results by semantic similarity, temporal relevance, and access frequency. Uses embedding-based similarity matching combined with recency weighting and usage statistics to surface the most contextually relevant memories. Supports both exact keyword matching and fuzzy semantic search, with configurable ranking algorithms to balance freshness vs. relevance.

Provides native integration adapters for LangChain and CrewAI agents, allowing them to use AgentRecall as a drop-in memory backend. Implements callback hooks that automatically capture agent actions, observations, and tool results into the memory palace without requiring manual instrumentation. Supports both LangChain's memory interface and CrewAI's agent state management, enabling agents to access memories through their native memory APIs.

Bidirectional sync between AgentRecall memory palace and Obsidian vault, treating Obsidian as a persistent knowledge graph backend. Exports memory palace rooms and relationships as Obsidian notes with wiki-link relationships, enabling human review and curation of agent memories. Supports importing Obsidian vault structure back into memory palace, allowing humans to seed agent memory with curated knowledge.

Automatically organizes memories into semantic rooms (conceptual domains) based on content analysis and user-defined room schemas. Uses clustering algorithms to group related memories and assign them to appropriate rooms, with support for hierarchical room structures (rooms within rooms). Enables agents to navigate memory by domain (e.g., 'user preferences', 'technical decisions', 'conversation history') rather than flat lists.

Provides a pluggable persistence layer abstraction that allows swapping storage backends (in-memory, file system, SQL database, vector database) without changing agent code. Implements a standard interface for memory read/write/delete operations with support for transactions and consistency guarantees. Includes reference implementations for common backends (JSON file, SQLite, PostgreSQL) and enables custom backend implementations.

Implements automatic memory pruning strategies to prevent unbounded growth of the memory graph. Uses configurable policies (LRU, semantic similarity deduplication, importance scoring) to identify and remove or consolidate redundant/stale memories. Supports manual pruning triggers and automatic background pruning based on memory size thresholds or age.

Intelligently selects and formats memories for inclusion in agent context windows, respecting token limits while maximizing relevance. Uses a multi-stage selection process: semantic filtering (retrieve top-k relevant memories), importance ranking (score by frequency/recency), and token-aware formatting (compress memories to fit within budget). Supports different formatting strategies (summarization, bullet points, full text) optimized for different context window sizes.

Logs all memory access operations (read, write, delete) with timestamps, agent identifiers, and access context. Enables audit trails for compliance and debugging, with support for filtering logs by agent, time range, or operation type. Provides analytics on memory access patterns (hot memories, access frequency, temporal trends) to identify which memories are most valuable.