mem-recall - Trellis Doc

mem-recall makes the AI invoke trellis mem whenever the user references past conversations, retrieve content from local Claude Code, Codex, OpenCode, and Pi session stores, and answer with session-id + verbatim quotation. Trigger phrases include last time, we discussed, what did I tell <Claude/Codex>, find ... last week, 上次, 之前, and other references to prior dialogue. Without the skill, the AI defaults to “I don’t have that context” or speculative answers. The skill’s frontmatter description field instructs the AI to run trellis mem in these cases, with a search → context two-step retrieval flow.

Prerequisites

Tool	Purpose	Required
Trellis CLI 0.6.0-beta.0+	Provides `trellis mem`	Required
Claude Code, Codex CLI, OpenCode, Pi	Source of past sessions	At least one

npm install -g @mindfoldhq/trellis@beta
trellis --version   # ≥ 0.6.0-beta.0

Install

npx skills add mindfold-ai/marketplace --skill mem-recall

Or install all marketplace skills:

npx skills add mindfold-ai/marketplace

Flag	Description
`-g`	Install globally to `~/.claude/skills/`
`-a claude-code`	Target a specific agent
`-y`	Non-interactive mode

Ask the AI which skills are available; mem-recall should appear in the list.

Trigger examples

No manual command needed. The following user messages trigger the skill:

last time how did we solve the wait_agent deadlock in #240?
which project did I discuss the plugin design in?
find what I told Claude about memory architecture last week
上次我们怎么处理 #240 的来着？

Retrieval flow

The skill instructs the AI to execute two steps. Step 1 — Candidate search

trellis mem search "<keyword>" [--cwd <project>] [--since <date>]

Multi-token AND search across cleaned dialogue. Returns ranked sessions. Score formula: (3 × user_hits + assistant_hits) / total_turns. User-turn hits are weighted ×3 because user wording reflects topic intent more strongly than AI elaboration. Step 2 — Content extraction

trellis mem context <session-id> --grep <keyword> --turns 3 --around 1

Returns the top-N hit turns plus surrounding context. Default character budget ≤6000, adjustable via --max-chars.

Cleaning before search

trellis mem strips the following before searching, so hits reflect actual dialogue:

prompt injections: <system-reminder>, <workflow-state>, <INSTRUCTIONS>, <environment_context>, etc.
Codex AGENTS.md preamble (first user message is dropped entirely)
tool calls and tool results (only text blocks retained)
pre-compaction history (Claude isCompactSummary / Codex compacted events replace older turns with a summary)

Data sources

Reads local files directly. No daemon, no index, no upload.

Platform	Storage
Claude Code	`~/.claude/projects/<sanitized-cwd>/*.jsonl`
Codex	`~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl`
OpenCode	`~/.local/share/opencode/storage/{session,message,part}/...`
Pi	`~/.pi/agent/sessions/<encoded-cwd>/<timestamp>_<id>.jsonl`

Out-of-scope use cases

Need	Tool
Search code	`Grep` / `Read`
Search commit history	`git log` / `gh`
Search current-project files/docs	`Read` / `Glob`

mem-recall is for AI conversation history only, not file or code search.

Performance

Scope	Time
Project-scoped 3-week search	~0.85s
Global, no time filter	~3s

Stateless. Each invocation cold-reads from disk; OS page cache absorbs IO so warm and cold runs perform similarly.

​Prerequisites

​Install

​Trigger examples

​Retrieval flow

​Cleaning before search

​Data sources

​Out-of-scope use cases

​Performance