START HERE
COMPARE // AI_MEMORY 01
Wyrm vs the alternatives
An honest look at how Wyrm compares to the common ways teams give AI a memory: managed memory services, framework memory, and plain retrieval-augmented generation. No trash-talk — each approach is good at something. This is about which fits your problem. Most tools here remember facts. Wyrm remembers facts andwhat failed, why each decision was made, and acts on them — so the comparison isn't just storage, it's memory versus memory-and-intelligence.
WHAT MAKES WYRM DIFFERENT 02
Wyrm's real differentiators
Three of these — negative learning, decision causality, and ground truths — are intelligence, not storage. They are why Wyrm is a memory and intelligence layer, not just a place to keep facts.
Local-first by default
Memory lives in a local SQLite database on your machine. No data leaves your environment unless you opt into sync. Most managed memory services are cloud-first — your context sits on someone else's server.
MCP-native
Wyrm is an MCP server, so it plugs straight into Claude, Copilot, Cursor, and Windsurf without a per-app SDK or glue code. Framework memory typically lives inside one framework's runtime.
Negative learning
Wyrm records what failed and blocks the AI from repeating a known counter-pattern. Most memory systems only store what to remember, not what to avoid.
Decision causality
Decisions are linked to the reasoning and downstream effects that depend on them, so a later change can surface what it invalidates. Plain vector stores have no concept of why something was decided.
No-LLM core
Core retrieval runs on full-text search plus embeddings — no extra LLM call in the hot path. That keeps recall fast, cheap, and deterministic. Approaches that route every read through an LLM add latency and cost.
Ground truths + scaffolds
Project ground truths and reasoning scaffolds are injected into every context brief, so the AI starts each session already oriented rather than rebuilding context from scratch.
FEATURE MATRIX 03
Side by side
Columns describe categories of approach, not named products — capabilities vary within each category, so "Varies" means an honest blanket claim cannot be made.
| Capability | Wyrm | Managed memory service | Framework memory | Plain RAG |
|---|---|---|---|---|
| Local-first storage (data stays on your machine) | Varies | Varies | ||
| MCP-native — works across multiple AI clients | Varies | |||
| Counter-pattern / negative learning | ||||
| Decision causality graph | ||||
| No extra LLM call in the retrieval path | Varies | Varies | ||
| Semantic + full-text hybrid recall | Varies | Varies | Varies | |
| Cross-project / cross-session memory | Varies | Varies | ||
| Self-hosted, no per-seat cloud dependency |
DECISION 04
When to choose what
Choose Wyrm when
You work across multiple AI clients, you want your context to stay on your own machine, and you want the AI to learn from failures — not just remember facts. Wyrm fits a developer or team that lives in Claude / Copilot / Cursor and wants one persistent, local memory behind all of them.
Choose a managed memory service when
You want zero infrastructure, are happy for memory to live in a vendor's cloud, and your application is built around that vendor's API. Managed services trade local control for hands-off operation, which is the right call for some teams.
Choose framework memory when
Your app is built entirely inside one framework and you only need conversational memory scoped to that runtime. Built-in framework memory is the path of least resistance when you are not leaving that ecosystem.
Choose plain RAG when
Your need is document retrieval, not evolving project memory — you want to answer questions over a fixed corpus. RAG is excellent at grounding answers in documents; it is not designed to track decisions, failures, or state over time.
QUESTIONS 05
Frequently asked questions
Is Wyrm a replacement for RAG?
Not exactly — they solve different problems. RAG retrieves passages from a fixed document corpus to ground an answer. Wyrm stores evolving project memory: decisions, failures, ground truths, and patterns that change over time. Wyrm uses retrieval internally (full-text plus embeddings), but its job is memory and learning, not document Q&A. Many teams run both.
Why does local-first matter for AI memory?
Project memory accumulates the most sensitive context you have — architecture, credentials-adjacent notes, decisions, and what went wrong. Keeping it in a local database means that context does not leave your environment by default. You can opt into encrypted sync, but the baseline is your machine, not a vendor's server.
What is counter-pattern blocking?
Wyrm records failures, not just successes. When the AI is about to repeat an approach that has already failed in this project, Wyrm can surface that prior failure and block the counter-pattern. Most memory systems only store what to remember; Wyrm also remembers what to avoid.
Does Wyrm need an LLM to retrieve memories?
No. Core recall runs on full-text search and embeddings, with no extra LLM call in the hot path. That keeps retrieval fast, cheap, and deterministic. Approaches that route every read through an LLM add latency and per-query cost.
Which AI assistants does Wyrm work with?
Wyrm is an MCP server, so it works with any MCP-capable client — Claude Desktop and Claude Code, GitHub Copilot, Cursor, Windsurf, and others. One memory layer sits behind all of them instead of a separate integration per tool.
One local memory & intelligence layer behind every AI client.
Local-first, MCP-native, learns from failures, tracks decisions. See what Wyrm stores, how it recalls, and how to install it in under a minute.
LOCAL_FIRST // MCP_NATIVE // NEGATIVE_LEARNING