About abagraph.

AI agents need memory that remembers when things were true and where they came from. A relational table forgets time, a vector index forgets provenance, and a folder of markdown files forgets everything the moment it grows. abagraph is not Postgres, not a vector DB, not a graph DB, and not a folder of markdown files. Agents need bitemporal, source-backed, compactable memory — not a patchwork stitched together from tools that were never built for them.

Every claim in abagraph is bitemporal (valid-time plus record-time) and source-backed, carrying provenance and calibrated confidence. When a new fact contradicts an old one, abagraph closes the prior fact instead of overwriting it — the history is the data, so you can always ask “what did I know at time T?” It is agent-native over MCP, exposing tools to digest, remember, recall, recall_as_of, whats_changed, build_context, and explain_fact. Automatic context-window compaction injects a confidence-ranked, token-budgeted packet before the agent runs, so prompts stay small and grounded.

abagraph defines a new category: the AIhouse, a warehouse-successor for the AI era. Where data warehouses and lakehouses were built for human analysts, the AIhouse is built to be the long-term memory and system-of-record for autonomous agents. As a managed platform it competes conceptually with the stitched-together vector DB plus Postgres plus markdown most teams assemble today.

abagraph is in early access. The open-source engine is licensed under Apache-2.0 and available on GitHub. Note that “abagraph” (pronounced “ABA-graph”) is unrelated to the “ABBAgraphs” cover band or any legacy academic tool of a similar name.