LoveDoLove/Veyra
Persistent, evidence-backed engineering memory plus AST-level code intelligence for DeepSeek Harness coding agents.
Veyra gives a DeepSeek Harness coding agent a persistent, evidence-backed engineering memory plus an optional structural code-intelligence layer — the 'Dual-Brain' architecture. Veyra Memory stores engineering lessons, root causes, decisions, constraints and causal chains (symptom → rootCause → remedy → verifiedOutcome) with evidence anchors, validation states (unverified → reviewed → verified) and authority levels (candidate → derived → canonical). It observes session events, distills an end-of-turn into candidate memories, and injects the most relevant records into the turn's context with a clear 'not repository truth' notice. Retrieval is hybrid: FTS5 BM25 plus symbol/path boosts, token-level semantic overlap, intent affinity and relationship-graph cohesion, with a transparent score breakdown. When the external codebase-memory-mcp binary is installed, seven cbm_* tools add call-graph tracing, AST-level symbol search, architecture summaries and code-grounded search; when it is absent, Code Intelligence degrades and the memory layer keeps working. Human surfaces include /veyra slash commands, a read-only Knowledge Observatory and a Network Graph at /veyra in DSH Web.
Install
dsh plugin --profile web add @lovedolove/veyraCopied verbatim from the README's 'Install' section. No placeholder is used. After install, restart DSH (or reload the profile) and Veyra activates automatically — the README states no API keys, no external services and no configuration are required to start using memory. The optional structural layer needs a separate external binary (codebase-memory-mcp) installed via its own upstream curl installer or pointed at with CBM_EXE.
Compatibility
Node >=22.5.0 <25.0.0 (uses node:sqlite, first shipped in Node 22.5). DSH >=0.1.2-rc.1 <0.3.0-0, verified on DSH 0.1.7-rc.2 (plugin boot, tools, /veyra + /veyra/graph) and 0.2.0-rc.1 (real compatibility gate + composed-host-service runtime); CI runs on Node 22 and 24. Installed per profile (web, tui, …), but the Settings section and the Network Graph require the web surface. It registers as a Cordis host plugin (src/plugin.mjs) with a node:sqlite + FTS5 store under the Veyra home (one DB per project).
Details
- Repo: LoveDoLove/Veyra
- Category: Memory
- Stars: 1
- Version: GitHub repo LoveDoLove/Veyra; npm package @lovedolove/veyra; MIT. The README states no explicit version number; it notes the project is on a pre-1.0 0.1.x line and its APIs and command surface may still change between minor versions.
FAQ
- Does Veyra write into my repository?
- No. The README states the repository stays authoritative and Veyra never writes into it; memory is stored in a node:sqlite + FTS5 store under the Veyra home (default ~/.dsh/veyra/), keyed by workspace path + git remote so two unrelated folders never share memory.
- Do I have to install codebase-memory-mcp?
- It is optional. The README (Code Intelligence) says the seven cbm_* tools and veyra_code_status require the external native daemon; without it Code Intelligence enters degraded mode and returns a structured error with installation guidance, while all Veyra Memory tools, the Observatory, slash commands and hybrid search continue working without interruption.
- What do the candidate / derived / canonical levels mean?
- Per the README (Memory rules): candidate = auto-captured observation (not auto-recalled); derived = remembered/learned lesson (auto-recalled); canonical = explicitly promoted truth (auto-recalled). Canonical is never assigned automatically — only /veyra promote <id> canonical on your explicit word.
Alternatives
Co-Engram/Co-Engram · flymysql/dsh-memory · highland0971/dsh-native-memory