Distillery 0.4.0 hardens memory API for agent systems
Distillery 0.4.0 stabilizes the MCP tool surface as a public contract, committing to backward compatibility and deterministic behavior for agent memory systems that depend on consistent retrieval and synthesis layers.
Score breakdown
Developers building agent systems can now depend on Distillery's memory layer as stable infrastructure; consistent tool contracts and deterministic behavior prevent downstream planners, evals, and shared knowledge bases from inheriting instability that would otherwise compound across the agent stack.
- 01Distillery 0.4.0 establishes the MCP tool surface as a public contract with guaranteed backward compatibility and major-version-only breaking changes
- 02The release prioritizes API stability over new features, with 60+ PRs focused on API surface hardening
- 03Distillery implements Karpathy's compile-and-query pattern using a single DuckDB table with hybrid BM25 plus vector search, not three-tier layered memory
Distillery 0.4.0 shifts focus from feature shipping to API hardening, establishing the MCP tool surface as a public contract. The release responds to a broader conversation about agent memory sparked by Karpathy's "LLM Wiki" concept, which inspired at least five projects (Knowledge Raven, Memoriki, OpenTrace KG-MCP, OptiVault, and Obsidian variants) to adopt a pattern of LLM-maintained wikis that synthesize knowledge over time rather than re-deriving context per query. Research into memory systems shows three-tier layered memory (fast index, episodes, raw transcripts) is now the default pattern in serious systems, with LongMemEval-S at 73% as a community baseline and MemPalace's ChromaDB stack posting 96.6% on LongMemEval R@5. Distillery builds on Karpathy's compile-and-query pattern using `/distill` writes, `/recall` queries, and `/pour` synthesizes, backed by a single DuckDB table with hybrid BM25 plus vector search and deduplication thresholds.
From 0.4.0 onward, tool names, parameter shapes, error codes, and response envelopes are guaranteed stable; breaking changes require major version bumps with deprecation windows.
The operational foundation of Distillery's memory layer rests on primitives that enable shared knowledge bases to admit error: every entry carries provenance (author, session ID, source), entries can be corrected without losing history, and entries can be marked expired or unreliable without deletion. From 0.4.0 onward, tool names, parameter shapes, error codes, and response envelopes are guaranteed stable; breaking changes require major version bumps with deprecation windows. This commitment ensures downstream planners, tools, and evals inherit deterministic infrastructure rather than drifting defaults that flood context windows or inconsistent error handling.
Key facts
- 01Distillery 0.4.0 establishes the MCP tool surface as a public contract with guaranteed backward compatibility and major-version-only breaking changes
- 02The release prioritizes API stability over new features, with 60+ PRs focused on API surface hardening
- 03Distillery implements Karpathy's compile-and-query pattern using a single DuckDB table with hybrid BM25 plus vector search, not three-tier layered memory
- 04Key operational primitives include provenance tracking (author, session ID, source), entry correction without history loss, and expiration marking without deletion
- 05Memory layer failures sit in the reasoning layer between retrieval and action, and when sessions end and volatile context evaporates, not primarily in retrieval itself