Context evaporation
Compaction, lost threads, and forgotten rationale force every new agent session to begin with a human reconstruction.
The persistent thinking layer
Work graph and durable memory for teams building with AI agents
Your work remembers
how it got better.
Fraimed gives people and AI agents one durable place for intent, execution, decisions, and the lessons proven by outcomes.
Connect and captureMCP-capable clientsClaude Code CLICodex CLIClaude.ai imports
Who it is for
Product and software teams using CLI and MCP-capable agents to plan, implement, and validate consequential work.
What you connect
A Fraimed workspace, supported agent clients, and opt-in capture hooks or conversation imports.
What you get
A shared Work Graph, a browsable project-routed record, and durable context agents can retrieve through MCP.
What is Fraimed?
An AI-native source of truth for teams that build with agents.
Fraimed is not a project manager with an AI feature, an agent orchestrator with tickets, or a chat wrapper with memory. It is one cohesive system where work, knowledge, agent execution, quality measurement, and durable memory share the same data model.
Work organization
Roadmap opportunities become structured, executable Frames with explicit validation and traceable attempts.
Persistent memory
Specs, decisions, references, conversations, and mementos survive sessions, agents, and team rotations.
Agent execution
MCP lets agents read governed context and return sessions, progress, evidence, and outcomes without copy-paste handoffs.
Quality feedback
Outcomes show whether work landed cleanly and grade the guidance that contributed to it.
The founding mission
Stop making people carry the context between machines.
AI-augmented work has a memory problem. Chat sessions compact and lose context. Specs live in one tool, decisions in another, and code in a third. The agent doing the next task cannot see what the last agent learned. The human becomes a router—re-explaining decisions and ferrying context between systems.
The fundamental unit of value is not the entire conversation. It is the consequential insight, decision, constraint, and pattern inside it. Fraimed captures those moments as structured memory, ties them to the work they shaped, and lets real outcomes determine whether they should guide the future.
The work graph is the substrate. Memory makes it persistent. Outcomes make it measurable. The feedback loop is the product.
Why it exists
Three failures traditional work tools were not built to solve.
Intent drift
Specs, tickets, code, and validation diverge across tools until nobody can prove whether the shipped work still matches the original need.
Invisible quality
Teams can see that an agent produced code, but not whether it succeeded first try, self-corrected, failed review, or shipped a defect.
One continuous loop
Context that compounds instead of evaporating.
The roadmap and work graph preserve what you intended. Sessions show what happened. Outcomes decide what deserves to shape the next attempt.
01
Frame the work
Turn intent into a clear hierarchy of projects, decisions, executable work, and proof.
02
Let agents work
Every connected agent reads the same live context and writes progress back through MCP.
03
Keep what worked
Outcomes grade the lessons behind the work, so useful context floats and weak context sinks.
The Work Graph
More than a list of tasks. A live model of how work gets done.
Fraimed connects strategy, capabilities, executable work, attempts, sessions, validation, and outcomes. Humans can see the whole hierarchy; authorized agents can navigate and update the same graph directly through MCP.
StructurePanelFrameWorkstreamSessionOutcome
Frame work with intent
Break a product goal into Structures, Panels, and executable Frames without losing the reason the work exists.
Validate before closing
Give every Frame explicit, checkable conditions. Evidence and waivers remain attached to the lifecycle instead of disappearing in chat.
Track every attempt
Workstreams distinguish one implementation approach from another; Sessions preserve each focused stretch of human or agent work.
See status and drift
Filter and inspect planned, active, review, and completed work while Fraimed surfaces unmet validation and work drifting from its north star.
Connect code and launch
Link branches, pull requests, files, tests, progress notes, and deployment evidence to the exact work they satisfy.
Learn from outcomes
Record whether an attempt landed cleanly, self-corrected, failed review, or shipped a defect—and use that verdict to grade its guidance.
Agent logs and recall
Keep the trail. Digest the signal. Recall it when it matters.
With supported live hooks installed, Fraimed captures CLI conversations so the work does not vanish when a terminal closes or a model compacts its context. The original conversation remains viewable while a deterministic digest makes it navigable.
Ingest continuously
Stop hooks and scheduled capture preserve active CLI sessions without interrupting the agent, with durable retry when delivery is temporarily unavailable.
Route by project
Conversation segments are assigned to the project they actually discuss. Ambiguous material stays visibly unassigned instead of contaminating another project’s memory.
Browse the real log
Open a conversation to read its timestamped messages in sequence, inspect topics and highlights, and search across the workspace.
Turn signal into memory
Agents can promote consequential excerpts into mementos, decisions, and work references—then future sessions retrieve the relevant context through MCP.
How capture works
Live capture is opt-in and client-configured. Supported Claude Code and Codex CLI hooks send conversation messages, source details, and timestamps to the user's authorized Fraimed account; failed deliveries queue for retry. Claude.ai conversations can be imported. Configured sensitive patterns are redacted before persistence. Project routing uses available project evidence, and ambiguous material remains visibly unassigned. The browsable digest is deterministic and requires no model call; promoting an excerpt into durable semantic memory is a separate action.