Platform Overview

memsearch provides plugins for 5 AI coding agent platforms. All plugins share the same core architecture: capture conversations to markdown, index with Milvus, recall via semantic search.

---

Comparison Table

Feature	Claude Code	Codex	DeepSeek Harness	OpenClaw	OpenCode
Plugin type	Shell hooks	Shell hooks	Native ESM + web client	TS registerTool	TS npm plugin
Capture method	Stop hook (async)	Stop hook (async)	session/event turn end	agent_end hook	SQLite daemon
Summarization	claude -p --model haiku	codex exec	DSH headless agent	OpenClaw agent	opencode run
Recall mechanism	SKILL.md (context: fork)	SKILL.md	Native skill	memory_search tool	memory_search tool
L3 transcript format	Claude Code JSONL	Codex rollout JSONL	DSH session database	OpenClaw JSONL	OpenCode SQLite
Isolation	Per-project collection	Per-project collection	Per-project collection	Per-workspace collection	Per-project collection
Install method	Plugin marketplace	install.sh	dsh plugin add	openclaw plugins install --force + hook permissions	npm + opencode.json
Embedding default	ONNX bge-m3 (CPU)	ONNX bge-m3 (CPU)	ONNX bge-m3 (CPU)	ONNX bge-m3 (CPU)	ONNX bge-m3 (CPU)
API key required	No (ONNX default)	No (ONNX default)	No (ONNX default)	No (ONNX default)	No (ONNX default)

Each plugin keeps its current native summarizer when the plugin-specific provider setting is empty. Claude Code, Codex, OpenClaw, and OpenCode also accept native; DSH selects its headless-agent backend when the provider is unset. To override one plugin's native model, set plugins.<platform>.summarize.model, for example plugins.claude-code.summarize.model, plugins.codex.summarize.model, plugins.dsh.summarize.model, plugins.openclaw.summarize.model, or plugins.opencode.summarize.model. To route summarization through a memsearch-managed API provider, define [llm.providers.<name>] and set plugins.<platform>.summarize.provider to that name. These plugin settings do not fall back to llm.model.

---

Shared Architecture

All plugins follow the same capture-index-recall pattern:

graph TB
    subgraph "Capture"
        CONV[Agent conversation] --> SUM[LLM summarization]
        SUM --> MD["memory/YYYY-MM-DD.md"]
    end

    subgraph "Index"
        MD --> WATCH[memsearch watch/index]
        WATCH --> MIL[(Milvus)]
    end

    subgraph "Recall"
        Q[User question] --> SEARCH[memsearch search]
        SEARCH --> MIL
        MIL --> RES[Relevant memories]
        RES --> EXPAND[memsearch expand]
        EXPAND --> TRANSCRIPT[Transcript drill-down]
    end

    style MD fill:#2a3a5c,stroke:#e0976b,color:#a8b2c1
    style MIL fill:#2a3a5c,stroke:#6ba3d6,color:#a8b2c1

Three-Layer Progressive Disclosure

Every plugin supports the same three-layer recall model:

Layer	Command	What it returns
L1: Search	memsearch search	Top-K relevant chunk snippets
L2: Expand	memsearch expand	Full markdown section around a chunk
L3: Transcript	Platform-specific parser	Original conversation verbatim

Memory File Format

All plugins write to the same markdown format:

# 2026-03-25

## Session 14:30

### 14:30
<!-- session:abc123 turn:def456 transcript:/path/to/session.jsonl -->
- User asked about Redis caching configuration
- Agent implemented cache middleware with 5-minute TTL
- Added Prometheus counters for cache hit/miss metrics

Cross-Platform Memory Sharing

All plugins write standard markdown and derive collection names from the project directory using the same algorithm. Memories are automatically shared across platforms -- no manual configuration needed:

• Memories written in Claude Code are searchable from Codex, DSH, OpenClaw, or OpenCode
• Same project directory = same collection name = shared memories
• Different project directories, including linked worktrees, are naturally isolated
• Different directories share memory only when configured with the same explicit [milvus].collection

---

When to Use Which

Scenario	Recommended Platform
Primary Claude Code user	Claude Code plugin -- most mature, marketplace install
Codex user	Codex plugin -- shell hooks, similar to Claude Code
DeepSeek Harness user	DSH plugin -- native lifecycle events, skill recall, and web memory browser
OpenClaw agent development	OpenClaw plugin -- native TS integration, multi-agent isolation
OpenCode user	OpenCode plugin -- npm package, SQLite-native capture
Using multiple platforms	Install plugins on each -- they share the same memory backend

---

Prerequisites (all platforms)

• Python 3.10+
• memsearch installed: uv tool install "memsearch[onnx]" (or pip install "memsearch[onnx]")
• First-time ONNX model download: ~558 MB from HuggingFace Hub (cached after first run)