Agents

Memory for AI agents that keeps its history.

"Your agent remembers" is no longer a product. Every serious agent tool now ships memory of some kind, on by default, scoped to the project in front of it. That memory is a cache: it is overwritten when the fact changes, it belongs to one tool, and nobody can say afterwards what it believed on a given day or who was allowed to read it.

What survives the free baseline is three things that a cache cannot be: an estate above any one project or tool, a record that keeps its history, and walls that decide who sees what. That is what Yohanun gives an agent.

An estate, not a cache

Memory lives in the platform, not in a tool's folder. Each agent has its own; a guild shares house knowledge to every agent you run; the same memory answers in Claude Code, Codex and any MCP host.

History, kept

A wrong memory is superseded, a finished one closed; both stay readable with their lineage. Ask what the agent believed on a date. Grades say how much weight a memory carries; a judge proposes changes and a person ratifies.

Walls, and a record

Owners, compartments and classification on every memory, decided inside the query before any model is called. One audit row per read. Outcomes reported against the memory that informed them move its ranking.

For coding agents: one command

pip install yohanun
yohanun init --api-key sile_... --owner "Your Name"
cd ~/dev/myapp && yohanun agent add myapp --matter project:myapp --project-dir . --wire

That wires two hooks: every session opens with where you left off and what is still open, and every prompt is answered with up to three relevant memories in context, searched with your actual question. The agent stores decisions, gotchas and open tasks as it works, and distils the session at the end. Past sessions can be backfilled into memory, with their raw turns kept below the floor for when an exact error or path is needed.

Nine agents have run on this for months inside the company that builds it, through the same kit customers install. The field notes are in the Guilds field experiment; the product page is Guilds.

For agents you build: one API

Store a memory, search it, upload a document, ask a grounded question, report an outcome, act within a mandate and escalate above it. Memory-profile keys never call a model, so your agent keeps whatever model and framework it already has; chat-profile keys let the platform run the model and meter it, or run it under your own provider key.

Request an invite