Transactions
The database
ACID rows for applications. Oracle, Postgres.
The system of record for AI agents
Your data isn’tready for AI.
Warehouses were built for analysts. AI agents need a system of record — accountable for facts across time, not just during retrieval. abagraph is that record.
SCROLL
The shift
The system of record changed. Your stack did not.
Every era of computing grew a system of record shaped like its main workload. The next one is agents reasoning over what they know: millions of small, time aware, source backed reads and writes. Rows built for dashboards are the wrong shape for that. So is a flat vector index. Snowflake and Databricks won the analytics era. The AI era needs its own platform.
Analytics
The warehouse
Columnar scans for analysts. Snowflake.
Agents
The System of Record
Bitemporal, provenance-preserving facts for agents and auditors. abagraph.
context.build(“nasdaily”)
resolve 11/12 sources, ranked by relevance
embed query, gather closest matches
compact remove duplicates, drop low signal
packet ready, sources attached
source tokenscontext tokens
214,80031,200
Context Manager
Your context window, compacted automatically.
Give it a subject or a prompt: nasdaily, or “summarise Acme’s renewal risk.” abagraph gathers across your sources, ranks by relevance, resolves conflicts, and compacts to your token budget, on its own. Your model gets a bounded, source backed context packet where every fact is traceable and weak claims are left out. Lower token spend, far fewer hallucinations.
In plain words: instead of dumping a giant pile of data on your AI, abagraph hands it a short, trustworthy briefing on its own. It costs less and makes fewer mistakes.
Platform
Four planes. One system of record.
FACTORY ingests governed facts from authorized sources. STREAM propagates every change as a bitemporal event. FACT GRAPH stores and governs them with full provenance. ACT/PLAY retrieves exactly what the agent needs, under budget.
FACTORY
Connect authorized sources. Declare what to digest, ignore, or keep raw. Every change becomes a governed fact.
STREAM
The bitemporal log is your event backbone. Conflicts detected at write time. Replay any point in history.
FACT GRAPH
Bitemporal, provenance-preserving system of record. Disputed claims preserved side-by-side. Nothing silently deleted.
ACT / PLAY
Smallest-sufficient context packets under explicit budgets. One interface instead of a separate retrieval server per source.
Not another database
Your AI data does not belong in yesterday’s database.
Postgres. A SQL store. A graph DB. A key-value store. A warehouse retrofitted with AI. Or the worst, a folder of .md files pretending to be memory. None were built for how an agent remembers. abagraph is.
Postgres / SQLRows and joins for apps. No memory, no time, no trust.
Graph DBLinks without meaning or provenance, with vectors bolted on.
Key-valueFast, but blind. No relationships, no recall, no reasoning.
Vector DBSimilarity only. No time, no source, no graph, no governance.
Warehouse / LakehouseSnowflake and Databricks, built for analysts. AI bolted onto the wrong shape.
.md filesA folder of notes is not a memory system. No recall, no safety, no scale.
abagraphMemory, meaning, time, trust, and operations in one engine. Built for agents, from scratch.
Agent native
Built for agents, not dashboards.
Your agents do not write SQL and they do not read result sets. They call tools and consume context. abagraph speaks that interface natively over MCP, and rests on five foundations agents can trust.
Bitemporal truthWhat was true, and when you knew it.
Provenance and trustSource, confidence, knows when to say no.
Hybrid recallMeaning, patterns, links and time, fused.
Agent nativeMCP tools, bounded packets.
TenancyEach team’s data, separate and private.
Seven principles
Proof over adjectives.
You have seen the pitch decks. We would rather show a reproducible benchmark, a shipped surface, and the discipline to cut anything that does not earn its place. Seven principles, one for each node of the mark, that abagraph is built on and held to.
01
Proof over adjectives
Every claim rides on a reproducible benchmark or a shipped surface: command, commit, output file. We apply the rule to our own marketing.
02
Strengthen a foundation
Every feature must strengthen one of the five foundations: bitemporal truth, provenance, hybrid recall, the agent interface, or tenancy. If it strengthens none, we cut it.
03
Bitemporal by default
Two time axes are the substrate, not a feature flag. Updates are explicit supersessions, never silent overwrites.
04
Beachhead before broadside
Win agent memory completely before claiming the warehouse. Each step is gated on the prior step’s proof.
05
Managed, not your burden
A platform you run on, not a database you babysit. Scale, upgrades, and operations are ours. You ship agents.
06
Enterprise from day one
Isolation, access control, audit, and observability are not add-ons. They are in the box, because your customers depend on them.
07
One platform, not a patchwork
Your AI data should live in one trustworthy system, not stitched across a warehouse, a vector DB, a graph DB, and a folder of notes. Consolidating that patchwork is the whole point.
Time travel
Ask what you knew, and when you knew it.
Every fact records when it was true and when you learned it. as_of(T) rewinds the whole graph to any moment: what you believed on March 3rd, and what was actually true then. Your warehouse cannot answer that. abagraph answers it natively, which is exactly what an agent needs when it has to explain a decision after the fact.
In plain words: you can rewind time and see exactly what was known at any past moment. That is how you audit an AI’s decision later.
Questions, answered
The short version, for people and for machines.
abagraph is the system of record for what AI agents know and do. It stores facts with provenance, timestamps, and confidence attached, and serves them to agents over MCP. It unifies semantic, graph, relational, and temporal recall in one managed Rust engine.
abagraph is not a relational, key-value, graph or vector database. It is a bitemporal system of record for AI agents that brings meaning, relationships, time, and trust into one store. It replaces the vector-DB plus RAG glue that most teams stitch together for agent context retrieval.
abagraph has a trigger engine that fires webhooks, agent prompts, and notifications when fact predicates change. Agents run as durable, resumable kernel loops that can be inspected and resumed from the last persisted step.
Each tenant's data is isolated by namespace, access is scoped per agent key and role, and every change is recorded in an immutable bitemporal audit trail. Fact-level policies govern what each agent identity can read, write, or act on.
Yes. abagraph emits structured events for every significant write to the operations log. The bitemporal log itself is the audit trail — every fact records when it was true and when it was recorded.
Yes. abagraph is a managed data platform for AI agents. It runs as a managed service or inside your own cloud, with namespace isolation, access control, and observability included.
Wiki
Every term, in plain words.
New to this space? Here is the whole vocabulary in everyday language. Anywhere on this page, hover or tap an underlined word to see what it means. Or search them all here.
AI agent
Software that uses an AI model to act on its own: answering questions, looking things up, and completing tasks step by step.
Bitemporal
Keeping two timelines for every fact: when it was true in the real world, and when you learned it. That lets you ask what you knew at any past moment.
Provenance
A record of where each piece of information came from, so you and your AI can check it and trust it.
MCP
Model Context Protocol. A shared standard that lets AI agents call tools and fetch data, like a universal plug that AI tools agree on.
Vector search
Turning text into numbers so a computer can find things with the same meaning, not just the same words. Finding the closest matches is called kNN.
Token and token budget
AI models read text in small chunks called tokens. A token budget caps how much you send, which controls cost and speed.
Context
The information you give an AI for one task. The context window is the most it can read at once.
Compaction
Shrinking a large pile of information down to just the parts that matter and fit the AI’s limit. abagraph does this automatically.
Recall
Looking up the right information from memory when the AI needs it.
as_of(T)
A way to ask “what did we know at this exact moment in the past?” and get the answer exactly as it stood then.
Supersede
Replacing an old fact with a newer one while keeping the old one on record, instead of deleting it.
Hallucination
When an AI gives a confident answer that is actually wrong, often because the information it had was missing or stale.
Latency, milli and nano seconds
How long something takes to respond. A millisecond is a thousandth of a second. A nanosecond is a billionth of a second, far faster.
Covering index
A prebuilt shortcut that lets a database find data instantly instead of scanning everything.
Data warehouse
A large database built for human analysts to run reports and dashboards.
Lakehouse
A system for storing huge amounts of raw data, used mainly by data scientists.
Tenancy and namespace
Keeping each team’s or customer’s data separate and private inside one shared system.
System of record
The official, trusted source where the real answer about something lives.
DAG
A directed acyclic graph: a map of steps that always move forward and never loop back. It is how you describe a multi-step job so it runs in the right order.
Workflow
A defined sequence of steps that runs reliably, with retries and approvals, even if something fails partway through.
Hooks
Small bits of logic that run automatically when something happens, like a new fact arriving or two sources disagreeing.
Plugins
Add-ons that extend the system with new sources, tools, or outputs.
Skills
Reusable abilities you can give an agent and share across your team.
Monitoring and alerts
Watching key numbers like freshness and confidence, and getting notified the moment something looks wrong.
Logs and traces
Logs are a timestamped record of what happened. A trace follows one request through every step so you can see where the time went.
Early access
For companies putting agents in front of customers.
Early access is open to companies with large data estates deploying AI agents to their customers. If your agents are going to production in front of real users, we want to build with you. Partners get hands on support, direct input on the roadmap, and first access before launch.
For data and AI engineering teams shipping customer-facing agents at scale. We review every request.
Request received.If it is a fit, we will reach out from the team at abagraph soon.