Harbor
- Type
- Project
- Status
- Active project
One of our own tools
One brain for every agent
Tedd is the memory and workflow layer we built for our AI agents and use on client work. Codex and Claude Code share one private memory of each project: what one learns, the other can use, every change shows which agent made it, and people can follow every connection in Obsidian.
Two agents, one brain
Codex, from OpenAI, and Claude Code, from Anthropic, work from the same project memory. Each sees only what its task allows, and every change shows which agent made it.
Follow the connections
Tedd writes its memory out as linked notes in Obsidian, so people can see how a decision connects to the people, work and evidence around it. Pick a record.
Pick any record
What it means for your project
The same discipline on every task, from the first week of a project to years in.
Nothing learned is lost. Decisions, the reasons behind them and what is still open are kept with their evidence. Ask why something was built a certain way months later, and the answer comes with its source.
Any agent picks up where the last left off. Whichever agent takes the next task starts from what the last one learned, not from a blank page, so nothing gets explained twice.
Finished means checked. Every task ends with the result checked and a list of anything left to do, rather than “done” on trust.
Evidence, not orders. Anything remembered, like a line in an old email, counts as evidence. It is never treated as an instruction or as permission to act.
Also inside
The parts that make it safe to trust with real client work.
- Recall scoped to the taskA task recalls at most twelve records, and only from the project it is for.
- History keptA correction is saved beside the earlier version. Nothing is quietly overwritten, and nothing disappears just because it is old.
- Rechecked before relied onFacts that may have changed are checked again before Tedd acts on them or reports work as done.
- Work split across agentsWhen a job benefits from it, Tedd splits it across focused agents and brings the results back together.
- Read-only by defaultAgents read memory through read-only tools. Saving is a separate, reviewed step.
- One local databaseThe memory is a single database on Flare’s own machines, not a hosted memory service.
- Secrets kept outPasswords, keys and secret links are not allowed into memory, and are redacted if one slips into a note.
- Measured, not assumedChanges to how Tedd recalls are tested on whether answers stay correct, not just on how much they save.
Tested like everything in the lab
We tried a way of recalling that used much less context. It looked better on paper, then got answers wrong. So we kept the method that didn’t.
Kept
Recall by recency
97.9%of answers correct
0required facts missed
Dropped
Recall by text match and recency
91.7%of answers correct, using 58% less context
4required facts missed
366,000records in the largest scale test Tedd has passed. One synthetic benchmark on one Windows machine, not a promise about every setup.
The recall study: 24 paired synthetic tasks under one model configuration, each recalling at most twelve records. The full methods, results and limits are in Tedd’s research report and scale report.
Read the code on GitHub