SymbioPulse: MCP-Native Neural Memory for Codebases

SymbioPulse is an MCP-first context layer for AI coding agents. It builds a persistent project memory from directory fingerprints, task-to-file synapses, DNA constraints, and reusable skills, then exposes that memory directly to mainstream agent clients through the Model Context Protocol.

The developer should only need to load the MCP server. Indexing, protocol injection, context assembly, and learning happen through MCP tool calls.

Visual Overview

Core Loop

1. Environmental Pruning: Ignore runtime noise such as .git, .symbio, .cursor, virtual environments, caches, build outputs, and dependency folders.
2. Directory Fingerprinting: Scan source directories and cache structural snowflake fingerprints in .symbio/fingerprints.json.
3. Context Perception: Convert fingerprints into scent zones in .symbio/scents.json.
4. Dormant Anticipation: The MCP runtime initializes itself on demand and waits for the agent's task.
5. Resonance Matching: sym_sniff(intent) first checks O(1) memory, then narrows the search to high-scent zones and ranks resonant files.
6. Synaptic Bonding: sym_form_synapse(task, file_paths) binds exact intents, vocabulary antigens, and token memories to the files that solved the task.
7. Negative Feedback Regulation: sym_add_dna and sym_add_skill persist constraints and reusable implementation knowledge for future tasks.

MCP Tools

Tool	Purpose
sym_sniff(intent)	Primary context entry. Auto-initializes, refreshes stale indexes, returns target files, scent zones, fingerprints, DNA, skills, and feedback instructions.
sym_check_dna(file_path)	Returns architectural constraints that must be respected before editing a file.
sym_form_synapse(task, file_paths)	Records successful task-to-file bindings and strengthens file relations.
sym_add_skill(file_path, skill_summary)	Stores concise reusable knowledge about a file.
sym_add_dna(target, rule)	Stores durable constraints learned from mistakes or project rules.
sym_fetch_dna()	Returns all recorded DNA constraints.
sym_status()	Shows current memory, scent, fingerprint, and skill counts.
sym_initialize()	Forces initialization and protocol injection. Usually optional because tools self-initialize.
sym_reindex()	Forces a full project reindex.

Agent Support

SymbioPulse is not Cursor-specific. The MCP runtime writes lightweight protocol files for mainstream coding agents so each client receives the same behavioral contract:

Agent surface	Protocol file
Codex / OpenAI agents	AGENTS.md
Claude Code	CLAUDE.md
Gemini agents	GEMINI.md
Cursor	.cursor/rules/symbiopulse.mdc, .cursorrules
GitHub Copilot	.github/copilot-instructions.md
Windsurf	.windsurfrules
Cline	.clinerules

Each file points the agent to the same MCP-first workflow: sniff first, check DNA before edits, form synapses after successful work, and record reusable skills or durable constraints.

If a protocol file already exists in the user's project, SymbioPulse preserves the user's content and only appends or refreshes a marked managed block:

<!-- symbiopulse:managed:start -->
...
<!-- symbiopulse:managed:end -->

Content outside that block is never rewritten by protocol injection.

Runtime State

SymbioPulse writes project-local memory under .symbio/:

• scents.json: directory scent map.
• fingerprints.json: structural directory fingerprints.
• synapses.json: exact, antigen, and token task memories.
• relations.json: file-to-file affinity graph.
• dna.json: project constraints.
• skills.json: learned file summaries.

The MCP runtime also writes lightweight agent protocol files so compatible agents know to sniff first, check DNA before edits, and form synapses after successful work.

Architecture

1. Neural Core: .symbio state, synapses, DNA, relations, and fingerprints.
2. Olfactory Engine: source pruning, directory scanning, static symbol extraction, and optional LLM-enhanced scent keywords.
3. Resonator Engine: file-level resonance ranking inside scented zones.
4. Autonomous MCP Runtime: self-initialization, freshness checks, protocol injection, context assembly, and learning tools.

There is no separate user-facing workflow or background watcher. MCP is the product surface.

Installation

Requirements

• Python package runtime: >=3.8.
• Verified development environment: Python 3.13.13 on Windows.
• Recommended launcher for Cursor and other stdio MCP clients: uvx from uv.

Install uv on Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Restart Cursor after installing uv so the GUI process can see uvx on PATH.

Cursor / MCP Clients

{
  "mcpServers": {
    "symbiopulse": {
      "command": "uvx",
      "args": ["--from", "symbiopulse", "sym-mcp"]
    }
  }
}

If Cursor cannot find uvx, use the absolute executable path returned by:

Get-Command uvx

Local Development

pip install -e .
sym-mcp

Enable optional LLM-enhanced semantic scent generation only when needed:

pip install -e ".[semantic]"

The default package intentionally does not install litellm; static fingerprints and fallback keywords work without it, and first-run MCP startup stays lighter.

Smoke Tests In Cursor

After adding .cursor/mcp.json, ask Cursor to verify the server once:

Use the symbiopulse MCP server to run sym_status and show the raw result.

Expected result shape:

SymbioPulse Status:
- Synapses: 0
- DNA Rules: 0
- Scent Zones: ...
- Directory Fingerprints: ...
- Skills: 0

Then test automatic usage without mentioning SymbioPulse:

Find the authentication entry files in this project. Do not modify files.

Expected behavior: Cursor should run sym_sniff first, then continue with native search only if it needs more detail. After the answer, it should call sym_form_synapse with the files that mattered.

Validate project rules:

Record a project rule: page components must not access localStorage directly; use the storage service instead.

Then ask for a change that would violate it:

Add direct localStorage access in a page component.

Expected behavior: before editing, the agent should call sym_check_dna and avoid violating the recorded DNA rule.

Troubleshooting

'uvx' is not recognized as an internal or external command

• Install uv, restart Cursor, or set command to the absolute uvx.exe path.

Cursor logs dependency downloads as [error]

• Lines like Downloading pydantic-core, Downloading pywin32, or Downloaded pygments are often install progress printed on stderr, not a server failure. Wait for the first install to finish and avoid repeated reloads.

Cursor still injects the old protocol text

• Include --refresh in the uvx args once to force a clean tool environment:

{
  "mcpServers": {
    "symbiopulse": {
      "command": "uvx",
      "args": ["--refresh", "--from", "symbiopulse", "sym-mcp"]
    }
  }
}

• Run sym_initialize once so the managed protocol block is refreshed.
• Check that .cursor/rules/symbiopulse.mdc contains ## Mandatory Tool Use, not the old ## Required Workflow.

pip appears to hang and ends with KeyboardInterrupt

• The command was interrupted before installation completed. Re-run it inside the activated virtual environment:

python -m pip install --upgrade build twine