dsh-memgas

Most agent frameworks discard all context the moment a session ends, so every new conversation in the same project starts from zero. dsh-memgas solves this by letting the agent carry project-level conventions, architectural decisions, and hard-won lessons across sessions, with a background pipeline that continuously organises, merges, and fades memories over time.

Retrieval runs four parallel channels—lexical full-text (FTS5 for exact identifiers and error codes), dense vector similarity, multi-granularity entropy routing (deciding whether to surface a whole session or a single turn), and graph expansion via Personalized PageRank for multi-hop cross-session links—then fuses them with Reciprocal Rank Fusion. A baseline floor and automatic degradation guarantee that the worst-case quality never drops below standard retrieval. Memories keep evolving after insertion: association, reconciliation, reinforcement, decay, abstraction, and re-association processes run in the background so duplicates merge, conflicts are annotated, stale entries archive, and clusters of related facts distil into higher-level conventions.

Built for developers running long-horizon multi-session projects in dsh who want the agent to retain project context without manual prompting. Zero configuration, purely local SQLite storage scoped by git remote, no external API key required, no model download, no third-party data egress. An experimental local ONNX vector model can be enabled for semantic generalisation, and the same memory library is exposed via memgas-mcp for Claude Code, Codex, and other MCP hosts.