跪拜 Guibai
← All articles
Backend · Interviews · Algorithms

DeepSeek Harness Memory Plugins: Eight Options, Zero That Deliver Team-Grade Memory

By GoAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Adopting a memory plugin without designing team governance locks you into personal recall; the survey's concrete takeaway is that you will still hand-build promotion and ACL logic, so factor that in before picking a backend.

Summary

DeepSeek Harness (DSH) treats memory as a processing chain across the agent lifecycle, with hooks like agent/pre-step, session/event, and turn/end. The survey groups plugins into three working modes: auto-recall injects memory before each step, on-demand retrieval leaves the model to call a search tool, and local-auditable options store readable files with candidate and verified states. Among the eight plugins, MemOS local and dsh-auto-memory fit personal use, OpenViking and Hindsight target project or platform layers, and MemOS Cloud and memsearch split between hosted API and Markdown-plus-Milvus retrieval. The survey's practical conclusion: every plugin solves personal or project memory, but none has public evidence of automatically turning multiple people's experience into shared, trusted rules. Teams need a separate governance layer for evidence, permissions, conflict resolution, and promotion.

Takeaways
— DSH exposes lifecycle hooks (agent/pre-step, session/event, turn/end) where memory plugins capture, recall, and inject context.
— Plugins split into three modes: auto-recall, on-demand retrieval, and local auditable files.
— MemOS local runs MemoryCore and a Viewer inside the DSH Node.js process, no JSON-RPC bridge or sidecar.
— MemOS Cloud calls /search/memory before the first model step and /add/message after a completed round; includeToolMemory controls tool-call retention.
— OpenViking stores memories, knowledge, and skills in a context database addressed by viking:// URIs, with ACL and multi-tenant support in progress.
— Hindsight ingests repository history and injects project decisions and conventions at session start; supports cloud, self-hosted, and local daemon.
— memsearch writes summaries to .memsearch/memory/ as Markdown and uses Milvus for hybrid keyword-vector search; local ONNX embeddings work without a server.
— ReMe injects usage guidance and a read-only reme_search tool; its default HTTP service has no API-key auth and should not be exposed publicly.
— dsh-auto-memory stores memories in $DSH_HOME/memory/ with candidate, verified, disputed, and superseded states.
— Mnemosyne is a local SQLite memory backend, but the official README has no DSH adapter yet, so it is not installable as a DSH plugin today.
— No surveyed plugin demonstrates automatic promotion from personal memory to trusted team rules.
Conclusions

The gap between personal recall and team governance is not a retrieval problem; every plugin optimizes recall, but none ships a permission, evidence, and conflict-resolution layer.

Auto-recall is a trade-off, not a win: it lowers the model's decision cost but increases noise and requires stricter permission checks.

The survey's repeated warning that local storage 'verified' does not equal team trust suggests the community treats memory correctness as a social process, not a technical feature.

Concepts & terms
DeepSeek Harness (DSH)
A plugin-first agent harness where models, tools, sessions, context, and memory are all extended through plugins.
agent/pre-step
A lifecycle hook that runs before the current model step, where memory plugins inject recalled context.
session/event and turn/end
Lifecycle hooks where agents capture conversation and tool events to trigger summarization or writes.
candidate/verified states
A memory entry starts as candidate, then a human or automated check marks it verified; disputed and superseded stop it from being injected.
auto-recall vs on-demand retrieval
Auto-recall injects memory automatically before steps; on-demand gives the agent a tool to search when needed.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗