Show HN: Cont3xt – Infinite context window on laptop

Virtual Infinite Context for Agents and LLMs
An MCP server that maintains a continuous rolling context window, always surfacing the most relevant memories for the task at hand — while respecting your token budget.

This repository includes an MVP MCP stdio server suitable for integration with the Cline extension for VS Code. It exposes a minimal, SQLite-only toolset over stdio with zero external services required by default.

• Version: 0.1.1
• Transport: MCP stdio
• Default storage: SQLite at ./data/memory.db (configurable; example config uses ./data/novel_memory.db)
• Optional backends: Qdrant / Neo4j (disabled by default in this MVP)
• Tools implemented:
  • memory_upsert
  • search_memory
  • context_pack
  • health_ping

Notes:

• Tool names use underscores for compatibility with the Python server decorator. Dotted aliases (e.g., context.pack) can be added later.
• All tools return a standard JSON envelope as text content: { "ok": boolean, "data": {...}, "meta": { "duration_ms": number, "backend": "sqlite", "version": "0.1.1" } }

Requires Python 3.11+

This creates ./data/memory.db if it doesn’t exist (or the path specified by config/env).

This starts the MCP server over stdio.

Add an MCP server named cont3xt that launches the stdio server:

Example configuration shape (adapt for your Cline settings UI/JSON):

Notes:

• Ensure virtual-context-mcp is in PATH (pip install -e . creates the console script).
• The --config argument is optional; environment variables can override values (see Configuration).

All tools return a JSON envelope as text content.

Envelope:

Upsert memory for a session from either combined content or a user/assistant pair.

Arguments:

• session_id: string (required)
• content: string (optional; direct content)
• user_input: string (optional; used if content not provided)
• assistant_response: string (optional; used if content not provided)

Response (example):

Simple substring search with a recency-biased score over recent SQLite chunks.

Arguments:

• session_id: string (required)
• query: string (required)
• max_results: number (optional, default 10)

Response (example):

Packs a token-budgeted context using recent conversation chunks. Reserves a small buffer and includes the current input if provided.

Arguments:

• session_id: string (required)
• current_input: string (optional, default "")
• budget_tokens: number (optional, defaults to CONTEXT_MAX_TOKENS)

Response (example):

Notes:

• If nothing fits within the budget (after a fixed buffer of 500 tokens), the pack includes: "reason": "BUDGET_TOO_SMALL".
• Truncation policy: respect budget strictly; sections are assembled most-recent-first, then presented chronologically for readability.

Returns basic status and configuration fingerprint.

Arguments: none

Response (example):

1. In Cline, run a task like “Improve function X in file Y; add a unit test and make tests pass.”
2. Cline should:
   • Call search_memory(session_id, query="project goals")
   • Call context_pack(session_id, current_input="<file diff request>", budget_tokens=12000)
   • Generate plan → apply small diff → run tests
   • Call memory_upsert with (user_input, assistant_response) to persist the session outcome
3. Acceptance:
   • From a clean clone, pip install -e ., launch Cline, and complete a guided code edit PR in one go using the above flow.

Defaults are provided in code, optionally loaded from a YAML file (e.g., configs/novel_writing.yaml) and overridden by environment variables.

Environment variable overrides (examples):

• CONTEXT_MAX_TOKENS (default 12000)

• CONTEXT_PRESSURE_THRESHOLD (default 0.8)

• CONTEXT_RELIEF_PERCENTAGE (default 0.4)

• CONTEXT_CHUNK_SIZE (default 3200)

• CONTEXT_TOKEN_MODEL (default cl100k_base)

• DATABASE_SQLITE_PATH (default ./data/memory.db)

• DATABASE_QDRANT_URL

• DATABASE_QDRANT_COLLECTION

• DATABASE_NEO4J_URL

• DATABASE_NEO4J_USER

• DATABASE_NEO4J_PASSWORD

For the MVP, only SQLite is used. Qdrant / Neo4j are intentionally not required; future versions can enable them via env/config flags.

Truncation policy:

• A fixed buffer of 500 tokens is reserved.
• Packing order: include current_input (if provided) and then pack recent conversation chunks until the budget is met.
• If no sections can be included, the pack includes "reason": "BUDGET_TOO_SMALL".

Entry points:

• Console script: virtual-context-mcp → virtual_context_mcp.main:main
• MCP server class: virtual_context_mcp.server:VirtualContextMCPServer

Local logs: standard output (set PYTHONUNBUFFERED=1 for real-time logs). SQLite file: ./data/memory.db by default.

Run unit/integration tests as needed (note that legacy story-focused modules may not reflect the MVP surface area).

• MVP runs in a safe default mode: only persists to local SQLite.
• No filesystem ingestion or external network writes by default.

• Optional Qdrant/Neo4j integration behind env flags with graceful fallback.
• Aliased/dotted tool names for compatibility (e.g., context.pack).
• Richer retrieval (hybrid, deduplication), entity/graph memory, and pressure relief policies.
• DevContainer / Docker image for reproducible environments.

MIT © elevend0g