workspace_memory_dsh

When an agent works across multiple sessions, one of the hardest problems is context loss: a new session opens and the agent must rediscover where the task left off, which approaches were tried, and why certain decisions were made. workspace_memory_dsh is a session-memory plugin built for the DSH (cordis) architecture. Under each workspace it maintains a structured project_memory directory so agents navigate memory by semantic "covering propositions" instead of scanning an untyped file tree.

The v3 memory model classifies atomic propositions extracted from conversations into two layers. The upper layer groups propositions under persisted "big proposition" nodes (semantic classification). The lower layer is a vector network: nodes and funnel entries are embedded, nodes connect to their k nearest neighbours, and fine-grained entries can link to multiple higher-layer nodes via many-to-many entryLinks. Beneath each node sits a three-layer funnel (core facts, regular memory, process logs). Retrieval is hybrid—vector cosine plus lexical scoring merged by RRF—so a workspace without an embedding endpoint gracefully degrades to word-face recall without losing graph structure. The plugin also exposes 18 Remote API methods, provides client-side UI surfaces (memory button, preview panel, sidebar tree), and auto-migrates older workspace data.

It is important to note the scope: workspace_memory_dsh is designed as Session Memory, targeting tasks typically completed within a handful to a dozen sessions as well as text-production workflows. Its goal is to help an agent quickly recover "where the task left off, what has been tried, why, and what to do next" in a new session. It is not a large-scale knowledge base or a general-purpose long-term storage system. It is best suited for DSH host users who need cross-session task continuity for their agents.