v0.3.0

A hardcore memory & planning engine for AI Agents — graph-based knowledge storage, MCTS tree-of-thoughts reasoning, and hierarchical goal planning. Zero core dependencies, pure Python 3.10–3.14.

pip install agentic-os-core

Highlights

This release turns the engine from "looks complete" into "actually correct." The headline is a fix to the MCTS core, plus a wave of new retrieval and memory-consolidation capabilities that complete the original roadmap.

🐛 Core correctness fixes

• MCTS no longer degenerates into a single chain. Expansion previously attached only the first generator candidate, so the search collapsed into a line where UCB1 and exploration_weight had no effect and the best-scoring candidate could be discarded. Expansion now attaches all candidates and evaluates each node exactly once (halving evaluator/LLM cost). Applies to MCTS and AsyncMCTS.
• Multi-agent access control now real — unshared memories are no longer visible to other agents; share_with validates ownership.
• recompute_importance + forget no longer wipes the graph (PageRank scores are normalised to [0, 1]).
• Fixes to the recency timeline leak, subgraph() edge counting, summarize_topics idempotency, inverted-index growth, duplicate sub-goal ID collisions, and HTML-visualiser XSS escaping.

✨ New capabilities

• BM25 relevance ranking for keyword retrieval, with native CJK (Chinese/Japanese/Korean) bigram indexing.
• Hybrid retrieval (recall_hybrid) — Reciprocal Rank Fusion of BM25 + vector similarity; degrades gracefully to BM25 with no vector store.
• Configurable forgetting curves (temporal_decay / DecayCurve) — exponential / linear / hyperbolic / custom.
• Streaming event system (EventBus) — callback, blocking-stream, and async-stream consumption for real-time agent coordination.
• Community detection (detect_communities) + topic summarisation (summarize_topics) — cluster memories into topics and compress episodic detail into durable knowledge.
• Parallel plan execution (Executor.execute_parallel) — dependency-aware concurrent goal execution.
• Persistence restore (LongTermMemory.from_graph) — rebuilds the recency timeline after loading from storage.

🔬 Quality

• 423 tests, 96% coverage, mypy --strict clean, verified on Python 3.10–3.14.
• Property-based tests, 20k-node scale smoke tests, and concurrency regression tests that enforce the documented thread-safety guarantees.
• py.typed now shipped — downstream mypy sees the library's type hints.

See CHANGELOG.md for the full list.