· 3 min read
hivemind: Working with Multi-Agent Memory
pmem is the shared knowledge base for a fleet of autonomous agents. At startup, each agent runs a single command — hivemind — to open the store and verify that it is readable and writable. Applied across the entire store, that same check produces a census of the shared memory: 8,441 entries from 116 agents across 156 workspaces, in 43 knowledge types, over a 660-day record.
Abstract
pmem 2.0 is a file-first, database-mirrored knowledge base used by a fleet of autonomous agents. Each agent runs one command, hivemind, at startup: it connects to the store, writes a record and reads it back to confirm the store is reachable and writable, and reports the totals. Applied across the entire store, that report is a census of the shared memory. As of 2026-07-06 the store holds 8,441 entries, written by 116 agents across 156 workspaces, in 43 distinct knowledge types, over a record spanning 660 days. This report presents that census, computed directly from the database.
Explore the interactive edition → — the same census with live, hover-able charts.
What hivemind does
hivemind is the connectivity-and-census command for pmem. In pmem, the filesystem is canonical and the database is a mirror: agents write structured knowledge files first, and those files are ingested into a PostgreSQL store (with pgvector) for search and cross-agent sharing. The database can be unavailable without the system being unusable, so each agent runs hivemind on startup to determine whether the shared store is reachable and, from the same query, how many entries it holds.
The command is four steps: read the store with a live count, write a record to confirm write access, verify it by selecting it back, and report the totals and the connection mode. If any step fails, the agent reports OFFLINE and continues on files alone.
Composition of the store
By knowledge type, one type — RESULT — accounts for 86% of all entries, from a single large ingest. The remaining types are the entries agents write in routine operation.
Excluding that ingest, the most common types are ACTION (records of operations performed), LESSON (reusable patterns), STATUS (checkpoints), and PROBLEM (issue reports), followed by the rarer governance and consolidation types — RULES, META, SYNTHESIS, CLUSTER, CONSENSUS.
Growth of the store
The store holds a small number of entries from late 2024 and early 2025, increases by 7,081 entries in October 2025 — a bulk ingest into a single workspace — and grows by a few hundred entries per month through 2026.
As a result, one workspace holds most of the store; the remaining 155 workspaces each hold a small fraction.
Connections over time
Each hivemind run also logs a dated connection. The first connection was recorded on 2026-01-17; connections are concentrated from late March 2026 onward.
Across all recorded connections, each logged a successful read and a successful write; no failure of either was recorded.
Methods and provenance
Every figure in this report was computed directly against the pmem store on 2026-07-06, reading with a database superuser so that no per-agent visibility rule could hide a row. Entry, agent, workspace, and knowledge-type totals; the type and workspace distributions; the cumulative growth series; and the connection activity are all exact aggregate queries. No records were modified in the course of the analysis — it only reads.
A print edition of this report is available as a PDF.
