dsh-memory-s3

DeepSeek Harness memory has long been confined to a single local filesystem: swap machines, change a deployment, or spin up another DSH instance, and every accumulated user preference, project decision, and meaningful moment is lost. dsh-memory-s3 relocates the persistence endpoint to any S3-compatible object store (AWS S3, MinIO, Cloudflare R2, Alibaba Cloud OSS). Share one bucket key and every instance can read and extend the same living memory, eliminating the single-device ceiling entirely.

Core capabilities include a structured memory model with five entry types (preference, project, decision, history, moment) enriched with importance, tags, subject, timeline, linked references, and lock protection; pluggable embedder-driven vector semantic search with cosine top-k and metadata/keyword hybrid recall, optimized for Chinese text; a local cache projection that degrades to read-only when offline; a mandatory write-approval gate routed through the DSH approval seam; a three-chain audit trail (approval pairs, audit ledger, snapshots) that can reconstruct every write from session logs; photo and file attachments stored as immutable S3 objects with triple validation (extension whitelist, magic-byte sniffing, 20 MB size cap); and an automatic backlink index for querying which entries reference a given one. Tiered snapshot injection projects content in three layers (Bonds, Moments, Facts) with association markers and lock guards preventing accidental overwrites.

Ideal for developers and teams who need persistent cross-session memory across multiple devices or deployments: whether syncing between a local workstation and a cloud container, or sharing a single project context within a team, one bucket and credential setup is all it takes for every instance to participate in the same ever-growing memory.