FeedRecall

August 18, 2026 · View on GitHub

CI License Node.js MCP GitHub stars

Your feed is an inbox, not a knowledge base.

FeedRecall turns human-selected likes, bookmarks, saved links, repositories, videos, and articles into persistent project-aware memory for AI agents. It runs locally, keeps an evidence trail, and exposes the same memory to Codex, Claude Code, Cursor, and other MCP clients.

No X API. No cloud account. No telemetry. Your archive stays on your machine.

v0.2.0 is the current release: the local core, projects, and context packs ship today, while connectors and enrichment workflows are still growing.

FeedRecall local memory cockpit

Try it in one command

The interactive demo uses only synthetic records and creates its vault in a temporary directory. It never reads your browser, X account, cookies, or private files.

pnpm --config.ignore-scripts=true dlx https://github.com/Paoladev45/feedrecall/releases/latest/download/feedrecall.tgz demo

Open the URL printed by the command to inspect the inbox, evidence states, project relevance, timeline, and review queue. The demo automatically chooses the next free local port when 4173 is already in use. Press Ctrl+C to stop the local cockpit.

Requires Node.js 22.16+ and pnpm 11. The package comes from the latest GitHub release, not from npm.

Connect your AI agent

One command creates the local vault and installs a version-pinned MCP configuration:

# Codex, ChatGPT desktop, and the Codex IDE extension
pnpm --config.ignore-scripts=true dlx https://github.com/Paoladev45/feedrecall/releases/latest/download/feedrecall.tgz setup codex

# Claude Code
pnpm --config.ignore-scripts=true dlx https://github.com/Paoladev45/feedrecall/releases/latest/download/feedrecall.tgz setup claude

# Cursor
pnpm --config.ignore-scripts=true dlx https://github.com/Paoladev45/feedrecall/releases/latest/download/feedrecall.tgz setup cursor

The command disables dependency install scripts and uses the prebuilt package contents. The installer stores the exact release URL in the client configuration, so an existing setup does not silently move to a newer version. FeedRecall can also be connected to Claude Desktop, Cline, Gemini CLI, OpenCode, or any stdio-compatible MCP client; see the MCP client guide.

Why

Saving a discovery is only the beginning:

capture -> enrich -> classify -> connect to projects -> verify -> test -> adopt or reject

FeedRecall remembers both what a source claimed and what you or your agents later observed. Six months later, an agent can avoid recommending a tool that already failed your Windows test.

What makes it different

FeedRecall is not another generic "store a note, retrieve a note" memory server:

  • Human-curated input: your like or bookmark is the intent signal; the system does the organizing.
  • Project-aware memory: one discovery can be relevant to Roblox, agent tooling, or several projects at once.
  • Evidence lifecycle: claimed, observed, verified, tested, adopted, rejected, and replaced are distinct states.
  • Time-aware review: publication, capture, and last-seen dates stay separate; volatile offers can be reviewed without deleting durable concepts.
  • Agent-ready output: Codex, Claude Code, Cursor, and other MCP clients receive bounded, sourced context instead of an unfiltered archive.

The goal is simple: recover a useful discovery in seconds, then remember what happened when you actually tried it.

Develop locally

Clone the repository when you want to inspect or contribute to the code:

git clone https://github.com/Paoladev45/feedrecall.git
cd feedrecall
corepack enable
pnpm install --frozen-lockfile
pnpm run build
pnpm feedrecall init
pnpm feedrecall import examples/discoveries.json
pnpm feedrecall import-projects examples/projects.json
pnpm feedrecall search "Roblox MCP"
pnpm feedrecall recall "MCP memory" --project agent-memory
pnpm feedrecall context --project agent-memory --output context/agent-memory.md
pnpm feedrecall timeline --date-field published --group-by month
pnpm feedrecall serve

For the complete walkthrough, see the one-minute demo.

The canonical database is ~/.feedrecall/feedrecall.db, so local agents share one memory even when they run from different project folders. Set FEEDRECALL_HOME only when you intentionally want an isolated vault.

Optional local enrichment requires Ollama and a local model:

ollama pull qwen3:4b
pnpm feedrecall process --model qwen3:4b

Then connect this local checkout instead of the release package:

pnpm feedrecall install-client codex
pnpm feedrecall install-client claude
pnpm feedrecall install-client cursor

Agents can use memory_recall to recover a forgotten discovery and memory_context_pack to load a bounded project brief with sources, dates, and evidence status. memory_timeline groups the same memories by publication, capture, or observation date. The equivalent CLI commands are:

pnpm feedrecall recall "the MCP memory tool I saw last month" --project agent-memory
pnpm feedrecall context --project agent-memory --output context/agent-memory.md
pnpm feedrecall timeline --date-field first_seen --group-by week

See the MCP client guide for Codex, Claude Code, Cursor, Claude Desktop, Cline, Gemini CLI, OpenCode, and other stdio-compatible clients.

Reproducible benchmark

Run the benchmark locally with synthetic data:

pnpm benchmark

It builds the core, imports 1,000 deterministic records into a temporary vault, measures import, search, recall, and timeline latency, then reports the database size. The benchmark does not make network requests or read personal data. Results depend on your machine, so the command is the source of truth.

What v0.2 includes

  • One-command install from GitHub releases via pnpm dlx (never npm), shipping as feedrecall.tgz plus feedrecall.tgz.sha256.
  • Version-pinned MCP setup (setup codex|claude|cursor and install-client) that records the exact release URL so an existing client is not silently upgraded.
  • Public MCP launcher with a Windows cmd.exe shim (injection-hardened) alongside the direct pnpm … mcp form.
  • Local SQLite storage and full-text search.
  • Idempotent JSON and URL imports.
  • Processing, evidence, and decision lifecycles.
  • Projects with transparent relevance scores.
  • Timeline views grouped by day, week, or month using published, first-seen, and last-seen dates.
  • Recall forgotten discoveries with project-aware evidence ranking.
  • Generate compact Markdown context packs for individual projects.
  • Read/write MCP tools with explicit annotations.
  • Local dashboard and browser capture extension.
  • Optional local enrichment through Ollama; core features work without it.
  • Automated GitHub release workflow and CI on Ubuntu and Windows.

Capture without the X API

FeedRecall does not require the paid X API. Load the unpacked extension from extension/, open your Likes or Bookmarks page while already signed in, and export the posts visible in your browser. FeedRecall never asks for your X password and never stores browser cookies.

The extension captures only posts currently rendered in the page. Scroll, scan, export, then import the JSON. Repeating the flow is safe: stable source identifiers prevent duplicate memories while firstSeenAt and lastSeenAt preserve the timeline.

Public post URLs can also be enriched with local collectors such as gallery-dl and yt-dlp. Network sources are treated as untrusted data, never executable instructions.

Privacy

  • Local-first by default; no telemetry.
  • No remote model provider in the core application.
  • Tokens, cookies, and passwords are not accepted in imports.
  • The public repository contains only synthetic examples.
  • Destructive social cleanup is intentionally outside v0.2.0.

Current status

v0.2.0 already includes the local store, browser capture extension, timeline, recall, project context packs, MCP tools, and explainable freshness review. It does not claim to be an autonomous truth oracle: external release, pricing, or changelog verification is a planned enrichment step.

Roadmap

  • v0.1 Memory (done): import, enrich, classify, search, timeline, MCP.
  • v0.2 Projects (current): project objects, context packs, collections, stronger relevance.
  • v0.3 Intelligence: opportunities, experiments, decisions, agent feedback loop.
  • v1 Knowledge OS: discover, understand, suggest, test, implement, remember.

See docs/PRODUCT_PLAN.md for the product model and release plan, and docs/ROADMAP.md for acceptance criteria and boundaries.

Architecture

Browser / JSON / URL connectors
              |
              v
        SQLite + FTS5
              |
      +-------+--------+
      |                |
 Local cockpit     MCP server
                       |
          Codex / Claude Code / Cursor

ChatGPT desktop, Codex CLI, and the Codex IDE extension share the same local MCP configuration, as described in the official Codex MCP documentation. ChatGPT web requires a future remote, authenticated plugin and is intentionally outside this local-first release.

Contributing and security

See CONTRIBUTING.md and SECURITY.md. Never attach a real personal vault or browser export to a public issue.

If FeedRecall helps you recover one forgotten tool or decision, starring the repository and sharing a reproducible use case are the most useful ways to support it.

License

Apache-2.0