Operating mode
When you're operating across tools
Every session starts from zero. You re-explain context, re-prompt corrections, re-establish what your agent already knew.
One of three operational modes of the same person: someone building an operating system for their own AI agents. Building pipelines mode, Infrastructure debugging mode.
You enter this mode every time you open Claude, Cursor, ChatGPT, or Codex to get work done. Nothing your agent learns in one session is guaranteed to be there in the next, and when it gets something wrong, there's no way to correct it that sticks. You're the context janitor — the human sync layer between every tool. Neotoma gives you continuity so you can steer instead of drive.
Escaping
Context janitor — human sync layer between tools
Into
Operator with continuity — steering, not driving
Turn-by-turn prompting → review-and-steer
Tax you pay
Re-prompting, context re-establishment, manual cross-tool sync
What you get back
Attention, continuity, trust in their tools
Same question, different outcome
Without a state layer, agents return stale or wrong data. With Neotoma, every response reads from versioned, schema-bound state.
Tasks
without state layer
What are my open tasks?
No tasks found.
with state layer
What are my open tasks?
3 open tasks. Next due: submit proposal by Friday.
Task created in Claude, invisible in Cursor
You told Claude to track a deadline. Later you asked Cursor for open tasks. The deadline didn't exist because each tool maintains its own disposable context with no shared state.
People & contacts
without state layer
Email the latest draft to Priya.
Sent to priya@oldco.com.
with state layer
Email the latest draft to Priya.
Sent to priya@newco.io.
Stale contact, wrong email sent
You updated a contact's email in one conversation. The next session used the old address because provider memory silently compressed or discarded the correction.
Financial records
without state layer
Find the Whole Foods receipt from Feb 8.
No receipts found matching that query.
with state layer
Find the Whole Foods receipt from Feb 8.
Whole Foods, Feb 8 ($47.32). Stored from conversation on Feb 8.
Receipt stored, then lost
You shared a receipt in a chat session. Weeks later, you needed it for an expense report. The AI had no record of it; conversation-scoped memory doesn't persist documents.
Events & commitments
without state layer
Do I have anything due this week?
Nothing scheduled.
with state layer
Do I have anything due this week?
Follow up with Kenji re: proposal, due Thursday.
Commitment forgotten between sessions
You told your AI you'd follow up with a client by Thursday. By Wednesday, neither tool remembered; the commitment was locked in a prior session's expired context.
◆
Why this happens
×No persistent state across sessions; every AI conversation starts from zero
×Fragmented document sources scattered across email, drives, and screenshots
×Repetitive context-setting in every new AI interaction
×Lost commitments and forgotten action items between sessions
×Personal data (receipts, contacts, preferences) stored in provider memory with no control over retention or training use
Failure modes without a memory guarantee
Lost commitments across tools
Tool-to-tool context loss
Silent state drift over time
Weak correction loop; no way to fix what the agent got wrong
Personal data in opaque provider memory with no deletion control
Memory locked to one vendor's ecosystem
Every session starts from scratch
You explain the same project context, preferences, and constraints in every new conversation. Provider-side memory is conversation-scoped at best; it doesn't follow you across tools, and it silently drifts as models compress or discard context.
Commitments vanish between tools
You tell Claude to remind you about a deadline. Later you ask Cursor for your open tasks. The deadline doesn't exist because each tool maintains its own disposable context. Action items created in one session have no guarantee of surviving to the next.
No way to correct what the agent got wrong
When an AI tool extracts the wrong date, associates the wrong contact, or misidentifies an entity, there's no correction mechanism. You can't tell the system "that's wrong" in a way that persists. The mistake reappears next time.
Your personal data in someone else's memory
Your receipts, contacts, health information, and financial records live in provider-hosted memory. There is no transparency into retention, no guarantee against training use, and no delete button. The data most personal to you is stored in a system you don't control.
◆
AI needs
What you need from your AI tools, and what current tools don't provide.
- Persistent memory that survives session resets and tool switches
- A single source of truth accessible from Claude, Cursor, Codex, and ChatGPT simultaneously
- Automatic extraction of commitments, tasks, and preferences from conversations
- Corrections that actually stick: fix once, fixed everywhere
- Full provenance: every stored fact traces back to the conversation or document it came from
◆
How Neotoma solves this
Neotoma removes the tax you pay re-explaining your world to every tool. Every conversation, entity, and commitment persists as versioned state. Switch between Claude, Cursor, and Codex without losing context. Correct once and the correction sticks.
[
Cross-tool persistent state via MCP
Every agent connected to Neotoma reads from and writes to the same memory substrate. Store a task in Claude and retrieve it from Cursor. State is shared, not siloed.
](/cross-platform)
Automatic entity extraction every turn
The agent loop extracts people, tasks, events, preferences, and commitments from every conversation turn and persists them as versioned entities before responding.
[
Corrections that stick
Submit a correction once. It creates a new observation that supersedes the incorrect value. Same question, same answer, every time. The correction traces back to when and why it was made.
](/deterministic-state-evolution)[
Full conversation replay
Every conversation and turn is stored with provenance. Inspect what was known at any point in time. Diff state across versions.
](/versioned-history)
◆
What actually changes
Without persistent state, you're the driver on every turn. Every prompt carries the full weight of what came before because the system won't hold it for you.
With Neotoma underneath, the pattern shifts. The agent arrives at each session already knowing what it knew last time. Your role moves from composing detailed instructions to reviewing what the agent already knows and correcting when it's off.
Less typing, fewer prompts, shorter sessions that accomplish more. You stop thinking about whether the system remembers and start thinking about what you're actually trying to do. Not "let me re-explain my situation"; "here's what changed since yesterday."
◆
Data types for better remembrance
The entity types you'll store most often.
conversation
Persistent chat sessions with full turn history across tools
message
Individual conversation turns with role, content, and extracted entities
task
Commitments, reminders, and action items with status and deadlines
note
Captured thoughts, observations, and reference material
contact
People and their details (email, role, organization, preferences)
event
Calendar events, deadlines, and temporal commitments
preference
User preferences and configuration that persist across sessions
receipt
Purchase records, invoices, and expense tracking
◆
When you don't need this
For one-off questions, quick summaries, or single-document analysis, your AI tools already work fine. Neotoma is for when you need what you told one tool to still be true when you open another, and for when you need to know that a correction you made actually stuck.
◆
Other modes
The same person operates in multiple modes. The tax differs; the architecture that removes it is the same.
[
Building pipelines mode
Pipeline building
Entity resolution by inference. Corrections that don't stick. Memory regressions you absorb because the architecture won't.
](/agentic-systems-builders)[
Infrastructure debugging mode
Infrastructure debugging
Two runs. Same inputs. Different state. No replay, no diff, no explanation.
](/ai-infrastructure-engineers)
◆
In operating mode, the tax is re-prompting, context re-establishment, and manual cross-tool sync. Neotoma removes that tax and gives you back the attention and continuity it was consuming. Same architecture removes the tax in every mode.
Built by someone who runs every workflow (email, finance, content, tasks) through the same agentic stack.