Show HN: Offline project memory for coding agents – no API key, no vector DB

Adaptive Agent Memory Research is an open, local-first project exploring how coding agents can retain useful knowledge across conversations without an API key, vector database, or cloud service. Its working memory tool records experiences in an append-only log and offers eight MCP tools, plus a command-line interface, for adding, searching, and managing project memories. The proposed design separates episodic records from knowledge consolidated into semantic, procedural, and failure memories. Its central question is whether future usefulness can guide what gets retained better than similarity to a query or frequency of retrieval. The repository stresses that the tool is a research engine, not a finished product: the storage, controller, and retrieval work, but the proposed memory categories and utility-based selection remain hypotheses.

The project reports 74 registered experiments, with many blocked, invalidated, or retracted; none has independent review, so results are exploratory. Findings highlight practical risks: superseded facts are often nearly identical to their replacements, making metadata essential to avoid stale retrieval; separating when a fact was true from when it was recorded improved historical answers and prevented future-information leaks. A small multilingual test found strong Polish-query recall with a local embedder, though its setup was limited and safety was not assessed. An association graph reduced recall in some tests, and an earlier alarming poisoned-instruction result was withdrawn after a corpus flaw was found. The key benchmark - comparing adaptive memory with ordinary compaction and vector search under equal context limits - has not yet been completed; a deterministic test corpus is now available to support it.