One memory for the fleet
Functions, classes, interfaces, types, and endpoints become deterministic nodes connected by what calls what.
For engineering teams
Your coding agents are guessing. At team scale, that compounds.
Memtrace gives every agent and every engineer the same deterministic map of your codebase. Structure, history, and blast radius, shared across the fleet, running inside your own environment.
The problem
One codebase, many agents, no shared memory.
Your team runs a growing fleet of coding agents against a codebase no single engineer holds in full. Each agent rebuilds its understanding from scratch on every task. None of them sees how a change ripples through the rest of the system. You get confident output that breaks something three files away, and engineers who spend their review time catching it.
The memory layer
One structural graph for the whole fleet.
Every agent and every engineer queries the same graph and gets the same answer.
See what a change touches
Every agent and every reviewer can ask the same graph for upstream callers, downstream dependencies, and history.
No per-agent re-indexing
The graph is built once, kept fresh incrementally, and shared across the people and agents working the repo.
Scale
Built for codebases that do not fit in a context window.
Memtrace indexes and queries at the scale real teams operate. The graph is built once and kept fresh incrementally, so retrieval stays fast and accurate as the codebase grows.
13.6sDjango 3.3K-file HEAD graph
0.19msaverage lookup across 50K nodes
42.5msp95 incremental freshness on 50 edits
Sovereignty
Your code never leaves your environment.
For teams under IP or regulatory constraints, this is the line between a tool security can approve and one they block.
Parsing stays on your machines
Code structure is parsed and queried inside your environment rather than uploaded to a model.
The memory store is yours
Teams and Enterprise can run the shared MemDB where security, IP, and regulatory constraints require it.
Deterministic answers
Retrieval comes from a structural graph, so the same query returns the same answer every time.
How it works
Deterministic retrieval with no model in the loop.
Parse
Tree-sitter turns source into typed symbols without asking an LLM to infer the map.
Connect
Calls, imports, endpoints, and cross-repo edges form a deterministic graph.
Remember
Each symbol carries history, timelines, and change episodes across the repo.
Serve
MCP-native tools expose the graph to the agents your team already runs.
Who builds it
Built by a technical team in Copenhagen.
Memtrace is backed by Dansk Industri and built for teams that need agent memory without giving up control of their source.
Already in production
Heavy users run millions of agent calls a month against Memtrace graphs.
Pilot
Start with one real codebase.
Run Memtrace with a small group of engineers and measure the difference against how your agents work today.
Choose one real codebase and a small group of engineers.
Stand up local or self-hosted MemDB with your agent workflow.
Measure review time, broken-change catch rate, and context spent before expanding.
Teams from $29 per seat.Enterprise and self-hosted, custom.
Request accessQuestions a CTO asks
Plain answers before the pilot.
Where does our code go?
Nowhere. Execution is local, the store is self-hosted, and code is not sent to a model or to us.
Which languages and stacks?
Rust, Go, TypeScript, JavaScript, Python, Java, C, C++, C#, Swift, Kotlin, Ruby, PHP, Dart, Scala, and Perl through the grammars Memtrace ships.
Does it work with our agents?
Yes. Memtrace is MCP-native, so it connects to the coding agents and agent shells your team already runs.
How long to set up?
The public Django benchmark builds a queryable HEAD graph in 13.6 seconds. A team pilot measures the real number on your codebase and hardware.
Fleet memory