boomzikazita/dsh-memory
DSH memory plugin: RRF multi-lane recall / confirm-based writes / Fernet vault / session harvest
dsh-memory is a persistent memory system for DeepSeek Harness with five-way RRF recall fusing evidence from L1 rules (MEMORY.md / USER.md), L2 session FTS5, L3 semantic (optional Honcho/MCP), L4 wiki and L5 time-window sources into a ranked evidence ledger with per-item trust labels. Writes are confirm-before-write: memory_suggest enqueues a proposal and memory_confirm / memory_reject decide it, with budget checks, sensitive-pattern scanning and automatic trigger: tags guarding what lands in L1. A Fernet-authenticated encrypted vault (VAULT.md.enc) keeps secrets with plaintext appearing only as → vault:<ref> references, session transcription incrementally indexes dsh session logs into FTS5, and maintenance tooling covers health checks, git backups, hit auditing and asset de-duplication.
Install
dsh plugin --profile web add github:boomzikazita/dsh-memoryGitHub install per README (EN primary with zh summary line). npm dsh-memory 0.1.0 exists but carries no repository field — ownership cannot be verified, so the GitHub route is primary. Install: dsh plugin --profile web add github:boomzikazita/dsh-memory (Python scripts under scripts/ are executed on demand by the Node host).
Compatibility
DSH web profile; Node host with Python available for scripts/; optional Honcho/MCP semantic layer (L3) and wiki root (L4) are config extras.
Details
- Repo: boomzikazita/dsh-memory
- Category: Agent Capabilities
- Stars: 0
- Version: git github:boomzikazita/dsh-memory (npm dsh-memory 0.1.0 — repo field empty, not claimed)
- Last push: 2026-08-18
- First seen: 2026-08-18
Recent updates
No changelog section in README; feature list covers RRF recall, confirm-writes, Fernet vault, transcription and maintenance.
FAQ
- How does recall combine so many sources?
- Five-way RRF: L1 rules files, L2 session FTS5, L3 semantic (optional), L4 wiki and L5 time-window sources are fused into one ranked evidence ledger with per-item trust labels.
- Can the agent write memory directly?
- No — memory_suggest enqueues a proposal and memory_confirm / memory_reject decide it; writes are gated by budget checks and sensitive-pattern scanning.
- Where do secrets live?
- In a Fernet-authenticated encrypted store (VAULT.md.enc); plaintext appears only as → vault:<ref> references.
Alternatives
aqsk-BLG/dsh-memory · Towzai/dsh-memory