Willder

Willder is a context layer that syncs your apps, agents and teams, so that every agent on a team works from the same knowledge. The product is in early access and is built around a simple promise: better context, tighter control and smoother coordination for every agent on your team. It works in three steps. First, knowledge is saved once — a decision you made is kept for every agent. Second, agents already know it: the next agent starts with it and no one explains it again. Third, you choose who sees what, and access can be taken away. Willder is aimed at teams running multiple AI agents such as Claude Code and Codex CLI, and its engine is MIT-licensed and self-hostable. In the product demo, Claude Code saves the fact "Launch moves to Friday." to Willder with memory_ingest, and the reply says it is stored and readable now. The problem Willder addresses is that agents otherwise start from zero. Each agent redoes work another agent already did, rules get explained more than once, and each agent can see more than it should. Willder's team describes the core issue plainly: a system prompt that says "only discuss project X" is an instruction, not a boundary. A folder of markdown is a good memory for one agent, but the situation changes when several agents, run by several people, share it. Willder is positioned as the layer that carries decisions, rules and facts between agents and keeps them under a single set of access rules, so coordination does not depend on a human retelling context to each new agent session. Shared knowledge is the first pillar. When an agent saves what it learned, the next agent starts with it. The demo shows Codex asking when the launch is, and its memory_query reading "Launch moves to Friday." together with the agent that wrote it, on its next call. Sharing is scoped: only agents you shared a fact with receive it, and shared agents only. This means a decision written once is available to the agents that need it, rather than being restated in every prompt or copied between tools. Rules work the same way: write a rule once, and every agent you share it with gets it. The content frames this as agents picking up each other's work rather than redoing it. Access control is the second pillar, and it is enforced with per-agent tokens. Each agent gets its own scoped, revocable token, limited to what you share with it. An agent reads only what its person may see, and every read and write is logged with who did it. Taking access away is immediate for new requests: the demo shows a token being revoked, its next call receiving HTTP 401 "Invalid or unknown capability token," and the refusal being written to the audit log. Likewise, an agent on a Marketing token asking the same question gets "No facts found." Willder is explicit about the limit: revoking access cannot erase information already received, and an answer already given cannot be taken back. Permissions are decided in the database rather than in the prompt. Consistency between agents is the third pillar. Willder versions facts, refuses stale writes, and tells each agent what changed, so there are no silent overwrites between agents. This covers the case of two agents claiming one task at once and the case of two agents writing the same fact at once; the later agent is told. In the published test, 30 such conflict cases across 12 read paths were checked and every one was caught. Deletion is also handled: delete data once and it leaves every live store within a minute, though backups are not covered yet. Unplugging an app stops its data from showing. Overall, Willder brings three things under one roof: your apps, your agents and your team. Connectors bring apps in under the same rules, with GitHub, Google Drive, Slack, Figma and PostHog available and Gmail, Notion, Linear, Jira and Google Calendar planned. On the agent side, Claude and Codex are supported, with Hermes, OpenClaw, Grok, Cursor, Gemini, GitHub Copilot, Windsurf and Perplexity planned. Agents run end to end with Claude Code, Codex CLI, Goose, LangChain / LangGraph, OpenAI Agents SDK and Vercel AI SDK. On the team side, a directory of people with roles such as QA, Engineer, Design, PM, Sales, Support, Data, Ops, Legal and Marketing is used to decide who can read a source: bring a source in and choose who can read it. Willder's engineering posts describe using vector RAG, graph RAG and a database together, because none of them alone decides who may see a result, which version is current, and where a deletion has to reach. Willder publishes measured results from its live product. With six agents writing, a new fact reached the other agents in a median of 538 milliseconds under normal load; the median stayed under one second in all three load conditions, and 95% arrived within 1.6 seconds in the slowest. In an access test, 75 agent reads were checked against each person's permissions, including after access was removed, and none went past them, while all 30 conflict cases were caught. On the GateMem benchmark, five of the 21 episodes (140 checks) scored 71.3 on the live product, with answers produced by GPT 5.4 and scored by GPT 4o as the judge as in the GateMem paper; private details leaked in 6.7% of privacy checks and deleted ones in 4.4% of deletion checks, and a Claude judge scored the same answers 76.5. A separate projection based on 48 saved agent sessions replayed under the current design estimates a median of 19.8% fewer tokens per task, likely between 6.7% and 26.2%, if agents take the same steps as before — a figure that is projected, not yet measured live. The benefits follow from those mechanics: agents stop redoing each other's work, no one explains the rules twice, and each agent sees only what it should. Teams get a record of who read or changed what, which supports review and accountability. Because reads are never charged and data is never locked in — read, export and delete always work — teams can try Willder on their own problem without committing their information to a closed store. Deletion propagating to every live store within a minute supports data-handling expectations, with the acknowledged gap that backups are not yet covered. Concrete use cases described in the content include an engineering team where Claude Code records a decision and Codex picks it up on its next call; a shared rule written once and inherited by every agent it is shared with; an access change where a revoked token's next call is refused and the refusal is logged; a marketing-scoped agent asking a question and correctly finding no facts; and two agents contending for the same task or the same fact, where the later one is told rather than silently overwriting. Broader scenarios include bringing GitHub, Google Drive, Slack, Figma and PostHog content under the same access rules, and self-hosting the MIT engine. On pricing, plans are published but marked as planned, not for sale yet. Free gives room to try Willder on your own problem with a monthly allowance and storage. Solo starts from $20 per month for one person and all their agents, with heavier use adding credits, and the first 100 accounts get Solo free for a year as founding members. Team starts from $25 per person per month with a minimum of two people, giving the team's agents one shared place with access rules and a record of reads; the allowance and storage grow with each person. Reads are never charged — the allowance counts writes, synced items and Willder's own agents — and if it did not help, the next month is free. Hosting it yourself is free under the MIT licence. Willder compares itself with Mem0, Zep, Supermemory and Hindsight on six capabilities and credits those tools with leading on recall benchmarks and number of connectors. In summary, Willder's value proposition is that agents stop starting from zero while staying inside boundaries you set. It combines shared, versioned knowledge with per-agent scoped tokens, logged reads and writes, conflict detection and deletion that reaches every live store, and it does so across the apps, agents and people a team already uses.