Agents don't just share memory. They notice what others are doing, inherit what others have learned, and understand complex projects together.
Every agent starts from zero. Context is lost between sessions, work is duplicated across a team, and the rules one agent discovered never reach the next. Coordination becomes a copy-paste of prompts.
Hivemind gives a team of agents one shared, append-only memory space — so they coordinate from the same context instead of re-discovering it.
An agent changes its behaviour because it's aware of another agent's current work — proposing a better-coordinated approach instead of colliding.
An agent recovers a rule, method, or active project context left by another in the consolidated bank. The handoff lives in the memory layer, not a manual prompt.
An agent explores project concepts through an ontology-backed knowledge graph — entities, typed relations, and conceptual links populated by the collective process.
The lattice H maps the system: live work signals on top, the consolidated bank at the consolidating centre, the ontology knowledge graph rooted at the base.
Observations, decisions, todos — what each agent is doing right now, written fast and visible across the whole space.
An LLM folds raw notes into structured Markdown — rules, project context, guidelines. The working memory other agents inherit from.
Ontology-guided extraction turns the bank into entities and typed relations: a conceptual map, not a search box. Protocol-derived, never the commit or audit source.
Raw notes refine into a consolidated bank, then into the long-term ontology graph: one continuous pipeline that distils information across short, mid and long. Across a team, sovereign instances keep the same space in sync, with no central server.
Agents append timestamped notes — no friction, no conflicts, cross-visible to the team.
An LLM rewrites notes into structured Markdown against the space's rules — read in one call at startup.
Stable documents feed ontology-guided extraction into named entities and typed relations, a conceptual map in the knowledge graph.
Every peer holds its own copy of the same space — same bank version, kept in step by signed sync events along each link. Each team keeps its own infrastructure, backups, and LLM credentials. No hub, no single point of failure.
Hardening the payment webhook path. Retry logic now keyed on idempotency tokens; auth refactor in flight under claude-code.
Consolidated from 12 notes · 4 min ago · bank v23
Spoken over MCP. Works with Claude, Cline, Codex, anything that speaks MCP.
health · whoami · about
create · rules · summary · export
note · read · search
read_all · consolidate · compact
connect · ingest · query · push · status
create · restore · download
tokens · scope · GC · purge
A minimal stack: S3 + an OpenAI-compatible LLM. No vendor lock-in, no local database.
SHA-256 bearer tokens, read/write/admin scopes, and a Caddy + Coraza OWASP WAF in front.
Every protocol event is journaled on S3. Provenance, ACKs, and commits are inspectable, not magic.
Apache-2.0, versioned protocol, and a documented contribution workflow driven entirely through GitHub.