arscontexta
www.arscontexta.org- Category
- AI Tools
- Rank
- No. 2578Tools index
Previous survey · No. 2461 ·
- Listed in
- #5 Give an agent memory
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Data, Retrieval & Knowledge
- Interfaces
- Agent Skill / Plugin
- Builder
- agenticnotetaking
- GitHub
- 3.5k stars
- Date
About
A memory infrastructure plugin for Claude Code that provides persistent agentic memory capabilities. It enables knowledge management and context retention across Claude Code sessions through an installable plugin.
What it does
Ars Contexta interviews you for two to four turns about your field and how you work, then writes a whole note vault for Claude Code to run. The output is plain markdown linked with wiki links, a set of slash commands that pull insights from sources and connect them to older notes, hooks that check structure on every write and commit to git, and a seven page manual. Every generated vault splits into three areas: the agent's own identity and goals, the knowledge notes, and operational state such as queues and session logs.
Why it's ranked here
Most note setups for coding agents are a template you copy. This one derives the layout from a conversation and records why: the setup maps your answers onto eight configuration dimensions with confidence scoring, checks fifteen required building blocks, and runs a pipeline smoke test before it finishes. It ships under the MIT license, keeps everything in local files with no database or cloud service, and bundles its own reasoning as a linked graph of research claims you can query directly. That is more design intent than a typical starter vault carries.
What's good
Fresh context per step is built in: the queue runner spawns a new subagent for each processing phase, so extraction, linking and verification do not share one crowded context window. Vocabulary adapts to your domain, so the notes folder can become claims or decisions while the three-way split stays fixed. Semantic search is optional; the vault works with ripgrep and topic maps alone. Three presets, including a research configuration with citation tracking, give a starting point if you would rather not derive from scratch.
Tradeoffs
Setup takes about twenty minutes and the README itself calls it token-intensive, plus two Claude Code restarts. It requires tree and ripgrep on the machine. The README lists four hooks, including one that saves session state when a session stops, but the shipped hook configuration only wires session start and the two write hooks, so that capture step is not visible there. Auto-commit fires on every write. The generated instructions forbid writing notes directly and route everything through an inbox pipeline, which is rigid if you want quick edits.
How to use it well
Fit it to someone who wants Claude Code to maintain a long-running body of notes, such as literature reviews or ongoing research, and is willing to follow a capture, extract, connect, verify routine. Run the setup once in a dedicated folder, not inside a code repository, since the auto-commit hook will commit every note write. Start with the matching preset if one exists. If you add semantic search, review the tool auto-approval list it suggests before accepting it. It is not a lightweight scratchpad for a single project.
Technical notes+
hooks/hooks.json registers a session-orient script on SessionStart (10s timeout) and, on PostToolUse matching Write, a write-validate script (5s) plus an auto-commit script (5s, async); there is no Stop entry, although README.md describes a Session Capture hook on Stop. .claude-plugin/plugin.json declares MIT and advertises 15 kernel primitives, 26 commands, 17 feature blocks and 3 presets. generators/claude-md.md composes the generated context file from 11 always-on and 6 conditional blocks with vocabulary substitution. agents/knowledge-guide.md defines a Sonnet subagent for note-quality advice. methodology/index.md and methodology/failure-modes.md are topic maps; several open questions in methodology/derivation-engine.md remain unanswered.
Observed
- License
- MIT (LICENSE and plugin manifest)
- Distribution
- Claude Code plugin installed from a plugin marketplace
- Interfaces
- Slash commands, Claude Code hooks, one subagent definition
- Storage
- Plain markdown files with wiki links; no database or cloud service
- Required tools
- Claude Code, tree, ripgrep; qmd optional for semantic search
- Hook events wired
- SessionStart and PostToolUse on Write; no Stop hook in the hook configuration
Read from README.md, LICENSE, .claude-plugin/plugin.json, methodology/index.md, agents/knowledge-guide.md, generators/claude-md.md, hooks/hooks.json, presets/research/starter/index.md, methodology/derivation-engine.md, methodology/failure-modes.md.
What it can do
Retain conversation context across Claude Code sessions
Previous conversation history and context data → Persistent memory state accessible in new sessions
Store and retrieve knowledge from previous interactions
Code snippets, discussions, and project information → Searchable knowledge base entries
Maintain project-specific memory contexts
Project files, discussions, and development decisions → Project-aware memory storage
Enable contextual code assistance based on history
Current code request plus stored context → Context-informed code suggestions and responses
Install as plugin into Claude Code environment
Plugin installation package → Integrated memory infrastructure within Claude Code
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.