Phloem

1# Phloem
2 
3**Your AI forgets everything when you close the tab. Phloem fixes that.**
4 
5[![Go](https://img.shields.io/badge/Go-1.24+-00ADD8?logo=go&logoColor=white)](https://go.dev)
6[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)
7[![Release](https://img.shields.io/github/v/release/CanopyHQ/phloem)](https://github.com/CanopyHQ/phloem/releases)
8[![MCP](https://img.shields.io/badge/MCP-compatible-green)](https://modelcontextprotocol.io)
9 
10Phloem is a local [MCP](https://modelcontextprotocol.io) server that gives your AI persistent memory across sessions. It works with **any tool that supports the Model Context Protocol** — an open standard. Today that includes Claude Code, VS Code, Cursor, Windsurf, Zed, Neovim, Cline, Warp, Continue, JetBrains, and more arriving every week.
11 
12You install it once. Every MCP-compatible AI tool on your machine gets memory.
13 
14```bash
15brew install canopyhq/tap/phloem   # or download binary from Releases
16phloem setup                        # auto-detects your tools
17```
18 
19---
20 
21## Why Not Just Use `.claude` Files or Markdown?
22 
23Every AI coding tool has some form of project context — `.claude/` directories, `CLAUDE.md`, `.cursorrules`, system prompts. They work. So why do you need Phloem?
24 
25**They're siloed.** Your `.claude` files don't help Cursor. Your `.cursorrules` don't help Claude Code. Switch tools and you start from zero. Phloem is one memory that works across all of them because MCP is a shared standard.
26 
27**They're static text.** You write rules by hand and hope the AI reads them. Phloem memories are created naturally during conversation, found via semantic search (not keyword matching), and linked to the code they reference.
28 
29**They can't track what changed.** You wrote "we use JWT with 15-minute expiry" in a markdown file. Then someone changed it to 30 minutes. The file still says 15. Phloem **citations** link memories to specific lines of code and **automatically decay confidence** when the code drifts. Your AI never confidently cites stale information.
30 
31**They don't capture *why*.** Markdown files tell the AI *what* to do. Phloem's **causal graph** captures the chain of decisions — *why* you chose JWT over sessions, what that caused you to change, and what would break if you reverted it.
32 
33---
34 
35## What Makes Phloem Different
36 
37### Semantic memory, not keyword matching
38 
39```
40You: "Add the logout endpoint"
41```
42 
43Phloem finds "JWT tokens expire in 15 minutes, refresh tokens in 7 days" — even though nothing in your query mentions JWT or tokens. It understands *meaning*, not just words.
44 
45### Citations that know when code changes
46 
47```
48Memory: "Rate limiter uses sliding window, 100 req/min per API key"
49  → src/middleware/rate_limit.go:42-67 (confidence: 0.95)
50```
51 
52Refactor that file? Confidence decays. Delete it? Citation marked invalid. Your AI adapts.
53 
54### Causal graphs, not flat lists
55 
56```
57"Switched from REST to gRPC"
58  → caused: "Updated service clients to protobuf"
59  → caused: "Added buf.gen.yaml to build pipeline"
60  → caused: "Rewrote integration tests for streaming"
61```
62 
63Ask "what would break if we reverted gRPC?" and get a real answer.
64 
65### Completely offline
66 
67Zero network requests. Not as a policy — there is no networking code in the binary. Your memories are a SQLite file on your disk. `phloem audit` proves it.
68 
69### Tool-agnostic by design
70 
71Phloem speaks MCP — the same open protocol that Claude Code, VS Code, Cursor, Zed, and others already support. Add a new AI tool? If it speaks MCP, it already has access to your full memory.
72 
73---
74 
75## Quick Start with Claude Code
76 
77```bash
78brew install canopyhq/tap/phloem
79phloem setup claude-code
80```
81 
82No restart needed. Start a session and your AI has memory.
83 
84For **all other supported tools** (VS Code, Cursor, Windsurf, Zed, Neovim, Cline, Warp, Continue, JetBrains), see the **[Supported Tools Guide](docs/SUPPORTED_TOOLS.md)**.
85 
86Or just run `phloem setup` — it auto-detects everything on your machine.
87 
88---
89 
90## Install
91 
92**macOS (Homebrew):**
93 
94```bash
95brew install canopyhq/tap/phloem
96phloem setup
97```
98 
99**Windows / Linux:**
100 
101Download the binary for your platform from [GitHub Releases](https://github.com/CanopyHQ/phloem/releases), extract it, add to PATH, then run `phloem setup`.
102 
103**From source:**
104 
105```bash
106git clone https://github.com/CanopyHQ/phloem.git && cd phloem
107CGO_ENABLED=1 go build -o phloem .
108./phloem setup
109```
110 
111Requires Go 1.24+ and a C compiler (for sqlite-vec/CGO).
112 
113---
114 
115## How It Works
116 
117**SQLite + sqlite-vec** — Everything in `~/.phloem/memories.db`. Vector embeddings power semantic search. No external services.
118 
119**Causal DAG** — Memories linked by cause and effect. Your AI traverses the graph to understand full chains of reasoning.
120 
121**Citation verification** — Memories attach to `file:line` ranges. When code drifts, confidence decays automatically.
122 
123**MCP Protocol** — JSON-RPC over stdio. No HTTP server, no ports, no network surface. Any MCP client connects instantly.
124 
125---
126 
127## Privacy
128 
129There is no networking code in the binary. Verify it:
130 
131```bash
132phloem audit
133sudo lsof -i -P | grep phloem    # should show nothing
134```
135 
136No accounts. No telemetry. No analytics. Delete anytime: `rm -rf ~/.phloem`. [Full details →](docs/PRIVACY.md)
137 
138---
139 
140## Contributing
141 
142We welcome contributions. See **[CONTRIBUTING.md](CONTRIBUTING.md)** for guidelines.
143 
144- [Report issues](https://github.com/CanopyHQ/phloem/issues)
145- [Discussions](https://github.com/CanopyHQ/phloem/discussions)
146 
147---
148 
149Apache 2.0 License. Built by [Canopy HQ LLC](https://canopyhq.io).
150