Recall Context

Proactive memory retrieval that pulls prior context relevant to the work about to happen. Reduces "starting from scratch" errors and prevents the agent from re-deriving conclusions already captured in mnemo.

When to Use

• Session start: load any preferences, decisions, or open questions

tied to the current cwd / project before the user types their first prompt.

• Before a significant decision: surface prior decisions in the same

area (e.g. database choice, lint rules, deployment target) so the agent does not contradict an earlier conclusion.

• When the user names a topic: e.g. "let's work on the auth flow" -

pull memories tagged or describing auth before proposing a plan.

• After a long context gap: if the conversation referenced earlier

decisions but the agent does not have them in working memory, recall them on demand.

Steps

1. Resolve query terms from the trigger:

• cwd: use the current working directory path + project name as the

query (e.g. mnemo-mcp or /c/Users/.../wet-mcp).

• recent: use the last 5-10 file paths the agent edited or read.
• <topic>: use the topic verbatim (the user's words).
• Default (no arg): combine cwd + last 3 file paths.

1. Search mnemo with context_type filtering when applicable:

   memory(action="search",
          query="<resolved query>",
          context_type=null,
          limit=10,
          include_archived=false)

• For decisions only: pass context_type="decision".
• For preferences only: pass context_type="preference".
• Without filter, results span all six context types.

1. Synthesize results into a 2-3 sentence summary grouped by type:

• decisions, preferences, facts, open tasks
• Present to the user as: "From prior sessions: ..."
• Include memory IDs for any item the user might want to update or

delete later.

1. Skip silently if mnemo is offline (tool errors), returns 0 results,

or only returns low-relevance matches (rerank_score < 0.3). Do not inject noise into the conversation.

Output Template

## Recalled context (<N> memories)

**Decisions** ([id1], [id2])
- ...

**Preferences** ([id3])
- ...

**Open tasks** ([id4])
- ...

(Recalled via mnemo:recall-context. Use memory(action="get", id=...) for full text.)

If 0 useful results: stay silent.

Quality Rules

• No injection on every prompt: this skill is invoked by the agent on

judgment, not auto-fired by a hook on every user message.

• Filter by context_type when intent is clear: searching all types is

noisy; restrict to decision/preference when the user is about to make a choice.

• Cite memory IDs: enables follow-up update/delete without re-search.
• Short summary > long dump: never paste raw memory contents - the

agent should already have full memory access via the memory tool if it needs detail.