CTXone

CTXone is persistent, searchable, accountable memory for AI agents. It eliminates context anxiety — the daily dread of re-explaining the same project to AI tools every morning.

The problem

Every AI session starts cold. Every conversation burns context tokens on re-onboarding. Architecture decisions made last Tuesday are gone by Thursday. CTXone gives agents a memory layer that survives sessions, restarts, and tool switches.

What CTXone provides

• remember — store facts, decisions, and reasoning with importance weighting
• recall — token-budgeted semantic retrieval; always returns a savings ratio
• prime — load markdown files as structured, idempotent memory
• Plans & tasks — shared plan primitive with state machine, proofs, blockers, and agent assignment
• Branches — every significant task gets a branch; decisions are auditable by task
• Sessions — end-of-session summaries written to the graph for the next agent
• Lens — web UI: dashboard, plans, sessions, memory tree, branches, diff

ASD code intelligence integration

CTXone Hub can spawn asd-serve on demand, giving agents five additional tools alongside memory:

• code_search — semantic symbol search
• code_read — full context for a symbol (signature, effects, ledger)
• callers_of / callees_of — call graph navigation
• code_repos — list registered repos

The savings model

Every recall response includes a ctx_savings_ratio — the ratio of tokens recalled vs. tokens that would have been spent re-establishing that context from scratch. CTXone Lens tracks cumulative savings per session and across the workspace.

Components

Component	Role
ctxone-hub	MCP server (42+ tools) + HTTP API
ctx	CLI — memory, plans, branches, taint, team ops
CTXone Lens	SvelteKit web UI
Python client	PyO3 bindings

Docs and source

• Docs: ctxone.com
• Source: gitlab.agentstatelabs.com/agentstategroup/ctxone
• License: BSL-1.1 → Apache 2.0 (4 years)