openclaw-plugin

LycheeMem OpenClaw Plugin

Purpose

This plugin is a thin adapter between OpenClaw and LycheeMem. It does not replace memory-core, does not claim plugins.slots.memory, and does not duplicate LycheeMem algorithms.

Collaboration model with OpenClaw:

• OpenClaw remains the host for the main reasoning loop, current-turn coordination, workspace guidance, and short-range conversational continuity.
• LycheeMem serves as the external structured long-term memory layer for cross-session recall, historical project background, entity relationships, and reusable procedures.
• Treat OpenClaw memory and LycheeMem as complementary layers, not competing memory owners.

Default plugin tool exposure:

• lychee_memory_smart_search (primary recall path, default lean response)
• lychee_memory_search (developer raw retrieval)
• lychee_memory_append_turn
• lychee_memory_consolidate

Use It For

• Historical facts the user mentioned earlier
• Long-running project context across sessions
• Entity and relationship recall
• Reusing procedural skills or workflows from prior work
• Compressing verbose retrieval results into a shorter background_context when needed

Do Not Use It For

• Workspace rules already covered by MEMORY.md or memory/*.md
• Stable preferences already maintained in memory-core
• Replacing OpenClaw's built-in memory owner
• Re-answering the current turn with a second memory system when OpenClaw already has enough local context

Coordination Rules

• Prefer OpenClaw's built-in memory and workspace context for same-session continuity, immediate local preferences, and repository-bound instructions.
• Prefer LycheeMem when the user is asking for longer-horizon context such as "上次这个项目怎么做的", "这个项目长期背景是什么", or "之前沉淀过哪些规则/关系/流程".
• Do not perform duplicate recall for the same question by calling both OpenClaw memory search and LycheeMem retrieval in the same turn unless the user explicitly wants a comparison.
• When LycheeMem returns a useful background_context, treat it as supplemental long-term context injected into OpenClaw's reasoning loop, not as a replacement for host memory.
• When OpenClaw already has enough local context to answer well, avoid unnecessary LycheeMem calls.

Trigger Guidance

• Prefer lychee_memory_smart_search for recall questions such as "上次怎么处理的", "用户之前提过什么", "这个项目长期背景是什么". Treat it as the default recall path.
• Let lychee_memory_smart_search use minimal output by default so the agent mainly receives synthesized background_context instead of verbose retrieval details.
• Use lychee_memory_search only during development or debugging when you explicitly want the raw retrieval payload.
• When this plugin runs inside OpenClaw with host lifecycle integration enabled, assume the host usually mirrors natural-language user and assistant turns into LycheeMem automatically.
• In that host-integrated mode, do not manually call lychee_memory_append_turn from the model during normal operation, because it would duplicate the host-managed transcript mirror.
• If host lifecycle integration is unavailable, disabled, or you are debugging a non-standard flow, call lychee_memory_append_turn manually after each completed dialogue turn so the transcript can later be consolidated.
• Do not append raw tool invocations, tool arguments, tool outputs, scratchpad text, or other orchestration-only traces unless the user explicitly wants those artifacts stored as memory.
• Do not call OpenClaw memory-core search and lychee_memory_search for the same recall problem in the same turn.
• When this plugin runs inside OpenClaw with host lifecycle integration enabled, assume /new, /reset, and /stop boundaries may trigger lychee_memory_consolidate automatically with background=true.
• Important long-term signals such as explicit memory requests, defaults, stable preferences, rules, and project standards may also trigger proactive background consolidation before the next reset boundary.
• Use lychee_memory_consolidate manually at wrap-up when host automation is unavailable, disabled, or you are debugging explicit persistence behavior.
• Even in host-integrated mode, it is acceptable for the model to call lychee_memory_consolidate when it intentionally wants to persist important new long-term knowledge early. The model should still avoid manual lychee_memory_append_turn in that case.

Recommended Pattern

The intended pattern is:

1. let OpenClaw evaluate whether its local memory, workspace instructions, and current-turn context are already sufficient
2. if longer-horizon recall is needed, call lychee_memory_smart_search with its default lean response
3. inject the returned background_context into the main reasoning context as supplemental long-term memory
4. answer in OpenClaw's normal reasoning loop
5. let the host lifecycle adapter mirror the natural-language user turn and assistant turn automatically when available; otherwise call lychee_memory_append_turn manually using the same session_id
6. do not append tool-call metadata or raw tool outputs by default
7. let host lifecycle boundaries trigger background consolidation when available, and allow proactive consolidation when the turn clearly introduces durable long-term knowledge; otherwise call lychee_memory_consolidate manually only if new memory-worthy information appeared in the mirrored natural-language turns
8. in host-integrated mode, if you choose to call lychee_memory_consolidate manually, do not precede it with extra model-driven lychee_memory_append_turn calls unless you are explicitly debugging transcript mirroring

Developer debugging path:

1. call lychee_memory_search
2. inspect the raw retrieval payload

This keeps OpenClaw in charge of the main reasoning loop while LycheeMem stays focused on long-term structured memory retrieval and persistence.