Context Infrastructure Reference Implementation
A context and memory blueprint for AI coding agents — persistent memory, personal rules, skills, and scheduled observations that make your assistant truly know you.
- Source repo
- grapeot/context-infrastructure
- Stars
- ★ 808
- Last updated
- 1d ago
- Primary language
- Python
- FA score
- 54/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer using Claude Code long-term who wants the assistant to remember technical preferences, project background, and communication style across sessions instead of re-explaining each time
- Not a fit if
- Users wanting out-of-the-box value; you must collect your own behavioral data over time
- English-only readers; primary docs and commentary are in Chinese
- Users unwilling to configure crontab jobs and .env environment variables
- Source review
- 54/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
This repository is the complete structure of a context infrastructure system the author has run for a year, positioned as a reference implementation rather than a turnkey tool. It is built around files and Markdown: AGENTS.md serves as the routing entry point for every session, the rules/ directory provides identity (SOUL.md), user profile (USER.md), and communication style (COMMUNICATION.md) templates, and contexts/memory/OBSERVATIONS.md carries the three-layer memory system. The periodic_jobs/ai_heartbeat/ folder contains observer.py (daily observation) and reflector.py (weekly reflection), two Python scripts scheduled via crontab to accumulate memory. The presentation layer — 43 decision axioms and 25+ skills — reflects the author's personal experience and is meant for reference only, while the reusable layer of templates and memory-system code can be adopted directly. The deployment boundary is a local repository plus scheduled jobs; heavier capabilities (search, email, documents) ship as separate public skill repos, keeping the main repo lightweight.
The flow starts by cloning the repo and opening the directory in Claude Code, OpenCode, or Cursor, where the agent reads AGENTS.md as its session routing table; filling in rules/USER.md with your preferences and background immediately personalizes agent behavior. Operationally, once crontab is configured, observer.py runs daily to write observations into OBSERVATIONS.md (the L1/L2 memory layers) and reflector.py runs weekly to produce reflections, following the SOP in periodic_jobs/ai_heartbeat/docs/KNOWLEDGE_BASE.md and the design in PRD.md. rules/skills/ holds 25+ reusable skills and rules/axioms/ holds 43 decision axioms as showcase samples. Scheduling guidance is in docs/CRONTAB.md; extended capabilities (web search, Google Docs, Google Maps, email/newsletter, PPTX, payment analytics, and more) are installed from the independent skill repos listed in docs/SKILL_ECOSYSTEM.md, with semantic search moved to grapeot/semantic-search-skill.
- A developer using Claude Code long-term who wants the assistant to remember technical preferences, project background, and communication style across sessions instead of re-explaining each time
- A user building a personal knowledge system who needs a scheduled daily-observation plus weekly-reflection pipeline (observer.py + reflector.py via crontab)
- A newcomer to AI coding tools who wants to fill in the rules/USER.md template and experience the personalized 'with context vs without context' difference within 5 minutes
- An engineer designing their own context system who needs a concrete reference for directory layout, data flow, and memory-layer design
- An advanced user extending an AI workflow who can selectively install independent skill repos from docs/SKILL_ECOSYSTEM.md for search, documents, email, PPTX, social media, and more
How do you install or deploy this agent?
- Clone the repo: git clone https://github.com/grapeot/context-infrastructure && cd context-infrastructure
- Open the directory with Claude Code / OpenCode / Cursor
- Configure environment variables based on .env.example (the README does not enumerate them)
- Open rules/USER.md and fill in your basic information — the highest-ROI step
- Full steps are in setup_guide.md; to enable the memory system, configure crontab entries for observer.py (daily) and reflector.py (weekly) per docs/CRONTAB.md
How do you use this agent?
- Each session starts with the agent reading the AGENTS.md routing table at the repo root
- Maintain your preferences in rules/USER.md and tune the agent's identity in rules/SOUL.md
- Let the observation scripts run on cron; memory accumulates in contexts/memory/OBSERVATIONS.md
- Survey reports, daily records, and thought reviews live in contexts/survey_sessions/, daily_records/, and thought_review/
- For more capabilities, install independent skill repos listed in docs/SKILL_ECOSYSTEM.md (web search, Google Docs, Google Maps, email/newsletter, PPTX, payment analytics, etc.)
- Note: true personalization requires collecting your own behavioral data from scratch; the showcased axioms and skills are reference only
What are this agent's strengths and limitations?
- A complete reference implementation refined over a year of real use — directory structure, data flow, and memory layering are all concretely inspectable, not just conceptual design
- Clean three-layer structure: template layer (USER.md, SOUL.md, COMMUNICATION.md) is fill-and-use, memory-system code (observer.py/reflector.py) is reusable, and personal axioms are clearly marked as reference-only
- Lightweight main repo; heavy capabilities (search, email, documents, payment analytics) ship as separately installable skill repos
- A full documentation chain — setup_guide.md, CRONTAB.md, PRD.md, KNOWLEDGE_BASE.md — covers configuration through memory SOP
- Core value depends on long-term collection of personal behavioral data; the README states plainly there is no shortcut, making cold-start cost high
- Limited out-of-the-box value: the 43 axioms and 25+ skills reflect the original author's perspective and cannot replace your own accumulation
- The memory system requires manually configuring crontab jobs, and the environment variables behind .env.example are not fully documented in the README
- The repository metadata lists the license as unknown while the README says MIT — verify before adopting
- Primary documentation and commentary are Chinese; English users must rely on the separate context-infrastructure-en repo
How does this agent compare with similar options?
The README splits semantic search into the standalone grapeot/semantic-search-skill repo and composes full capabilities from the multiple public skill repos listed in SKILL_ECOSYSTEM.md; the source names no other concrete competitors, so no comparison should be fabricated.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Context Infrastructure Reference Implementation This agent | 54 · Major gaps | Agent plugin / skillFree + model costs | ★ 808 | 1d ago | Python | Claude Code |
| Obsidian Mind | 79 · Good | CLIFree | ★ 5.4k | 5d ago | TypeScript | Codex · Claude Code |
| Keep the Why | 75 · Good | Agent plugin / skillFree | ★ 231 | 1d ago | Python | Codex · Claude Code |
| Mnemon — Persistent Memory Engine for AI Agents | 74 · Some gaps | CLIFree | ★ 652 | 3d ago | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence: README clearly describes data flow (three-layer memory, observations, cron jobs) with a .env.example template, giving adequate flow transparency. Deductions: observer.py/reflector.py run via cron, share_report publishes externally, and the VLM OCR path sends PDF content (potentially sensitive) as base64 to an external API — with no least-privilege design, user-confirmation gate, or sanitization guidance; no rollback or recovery path is documented; dependency security for external skill repos and .env secrets is template-only with no lockfile or audit note. MIT license and author background links justify a mid score for attribution.
Evidence: the test file covers CliError failure paths and a doctor command checking dependencies (requests/docling/lmstudio), showing deliberate failure and dependency handling; README structure and skill naming are internally consistent. Deductions: only one skill has visible tests; core scripts (observer/reflector) show no test evidence, and cron-job failure reporting is undocumented.
Evidence: README explicitly targets developers building their own context system, honestly positions itself as a reference implementation rather than a turnkey tool, and states capability boundaries ('display layer cannot be copied, data must be collected yourself'). Multi-environment support (Claude Code/OpenCode/Cursor) is stated. Deductions: cron trigger precision rests on a single docs file, and non-Unix platform fit is unaddressed.
Evidence: clean directory layout with a coherent three-layer split, complete install notes (Quick Start + setup_guide.md), honest known limitations, and explicit MIT license. Deductions: no CHANGELOG or version tags; maintenance responsibility is only the narrative of 'running for a year' with no update commitment; the 43 axioms and 25+ skills cited are not visible in reviewed files, so the examples score is conservative.
Evidence: outputs are structured markdown context plus runnable scripts; the 'with vs without context' quick-start comparison is a sensible design, and the frankness about required user investment is a plus. Deductions: core value depends on sustained user effort and cannot be verified statically; no cost hint for external VLM OCR API usage.
Evidence: most README claims trace to concrete files (directory tree, license, setup guide), and the README cleanly separates 'directly usable' from 'reference only', giving good fact/inference separation. Deductions: external linked skill repos are unverifiable from this repository, and numeric claims ('one year of operation', '43 axioms') rest solely on self-report.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- VLM OCR sends PDF content (potentially sensitive) as base64 to an external API; verify data boundaries and self-host the endpoint before use.
- observer/reflector run long-term via cron and write personal memory files; review script permissions and disk usage before deployment — the repo offers no rollback mechanism.
- Core value requires ongoing personal data collection; do not expect a turnkey tool. External skill repos are unverified within this repository — audit them separately before installing.
- share_report publishes reports externally; confirm sensitivity of content before enabling it.