imhamzakhan/context-flow-compressor

**The Context Fabric That Lets AI Agents Remember Everything—Without Re-Training**
Overview
Every AI conversation today suffers from a profound amnesia. You explain your project architecture, your coding style, your design philosophy—and the next session, the model greets you like a stranger. The token budget evaporates re-explaining context that should have been persistent.
**Semantic Memory Weaver** is not another chat wrapper or prompt manager. It is a **persistent context fabric** that weaves your conversations, decisions, and domain knowledge into a structured, traversable graph—local to your machine, invisible to the cloud, and instantly retrievable by any MCP-compatible AI tool.
Think of it as the **hippocampus for your AI stack**. Where other tools offer sticky notes, this project offers a neural architecture. Where others offer search history, this offers semantic recall. Where others cost you tokens to re-state context, this *pays you back* with every interaction.
The 2026 landscape of AI coding assistants is crowded, but almost all of them share a fatal flaw: they are **stateless by default**. This project inverts that assumption. Your AI doesn't need a bigger context window—it needs a *better memory*.
---
🧠 The Core Problem We Solve
The Token Tax
Every time you switch tasks, open a new session, or revisit a project after lunch, you pay a **token tax**. You re-describe your architecture, re-explain your conventions, re-state your constraints. Over a week, that tax amounts to 90% of your API spend being wasted on *repetition* rather than *generation*.
The Context Fragmentation
Your knowledge lives in:
- Scattered markdown notes
- Git commit messages
- Code comments
- Slack threads
- Pull request reviews
- Your own head
No single tool aggregates these into a queryable, structured memory. Semantic Memory Weaver **unifies these fragments** into a directed graph where each node is a concept (e.g., "rate limiter design") and each edge is a relationship (e.g., "depends on", "conflicts with", "supersedes").
The Privacy Paradox
Cloud-based memory tools require sending your proprietary context to third parties—a non-starter for many enterprises. This project is **local-first by design**. Your graph lives on your disk, encrypted at rest, and only your AI tool (via MCP) ever reads it. No cloud relay, no telemetry, no data exfiltration.
---
🚀 The Weaver's Approach: Graph-Based, Not Vector-Based
Most "memory" solutions for AI use vector embeddings—flattening your knowledge into a dense blob of numbers. That approach suffers from:
- **Loss of relational structure** ("Who said this?" is lost)
- **Poor update semantics** (revising one fact requires re-indexing everything)
- **Opaque retrieval** (you get "similar" text, not *the* answer)
Semantic Memory Weaver instead uses a **property graph model**:
Every node carries:
- **Content** (the actual text, code snippet, or decision record)
- **Provenance** (source file, timestamp, AI tool that produced it)
- **Confidence** (how validated this node is—human-authored vs. AI-inferred)
- **Access control** (which tools can read which subgraphs)
This graph is **traversable**. Your AI tool can ask: "Show me all decisions about authentication made after March 2026" and receive a subgraph, not a text dump.
---
✨ Feature Vault
1. **Local-First Graph Store** 🏠
Your memory is a TinkerPop-compatible graph database embedded in your project folder. No server, no network calls. The graph file is a single, portable binary (~2MB for 100k nodes) that you can version-control, back up, or move between machines.
2. **MCP Native Integration** 🔌
Implements the Model Context Protocol (MCP) out of the box. Any AI tool that supports MCP (Claude Desktop, Cursor, Windsurf, or your custom agent) can:
- READ a subgraph relevant to the current file being edited
- WRITE new nodes when a decision is made
- QUERY with natural language, resolved by a built-in semantic resolver
3. **Automatic Context Suggestion** 💡
When you open a file, the Weaver analyzes its imports, comments, and history, then **preemptively loads the relevant subgraph** into your AI tool's context—before you even ask. No manual prompting. No "remember that..." preamble.