lab report 02: shared memory and a way into the terminal
Period: August 1 to 4, 2025. Technical report on getting two AI agents to share information and giving them a memory that outlasted a session.
SUMMARY
By the start of August we were using two agents every day, Claude Code and Gemini CLI. They were good at different things and had no way of working together or remembering anything between sessions. In four days we tried three things: coordinating the two agents through a shared file, building a workspace with a proper memory system, and building a program that could get text into an agent's terminal from outside. All three worked to some degree, and all three shaped what came later.
COORDINATING THROUGH A FILE
Neither tool had any way to talk to the other, and there was no standard for it. The one thing both could reach was the filesystem, so on August 1 we set up a coordination file in a shared folder, and the two agents started coordinating through it.
Along the way we found out why Gemini had seemed to refuse certain tasks earlier. We'd been piping commands into it, and piped input skips the confirmation prompts Gemini uses to ask for permission. Its file tools were waiting for approval that never came. Once we stopped piping, it worked fine. That was useful to know, and it's still true of a lot of agent tools: non-interactive input and permission prompts don't mix.
DEVHUB AND ITS MEMORY
DevHub was an Electron desktop app meant to be the control center for every project. What mattered most in it was the memory design.
Memory was split into separate files by kind: sessions, projects, lessons learned and errors. Each agent had its own context file suited to its role. A master context file was read at the start of every session and kept current by a script that rebuilt it whenever the workspace changed. It used a compressed format to save tokens, and a short list of folders was marked off limits to the AI.
WHAT WORKED
The shared file. It worked the first day and every later system used some version of it.
The memory structure. Most of it is still how this lab's memory works today.
The off-limits folders. Simple, clear, and the start of a lot of later rules about what agents may touch.
WHAT DIDN'T
The shorthand. Compressing everything into abbreviations saved fewer tokens than we expected and made the model decode the file before it could use it. On top of that, the rebuild script walked every folder, including the one holding installed JavaScript packages, and wrote dozens of package names into the file meant to save space. Choosing what goes into memory saves far more than compressing it.
Agents keeping their own records. The abbreviation log was set up and never really used, and its entries still have placeholders where the dates should be. Agents are happy to design a process and inconsistent about following it. We ran into that again several times.
A WAY INTO THE TERMINAL
The harder problem was that an agent in a terminal only acts when someone types to it. Nothing else on the machine can reach it. If we wanted one agent to react to another, or a helper program to feed it context, something had to put text into the terminal from outside.
On August 4, between 4:21 and 4:54 in the morning, we wrote seven versions of a program called Symbiote, each taking a different route in. The early ones took the obvious routes. The later ones took much less obvious routes.
That last method was the best thing we found for this problem in the whole project, and far more dependable than anything that relied on the screen or keyboard. The session log ends on a failed attempt to find the console window, but the method did connect afterward.
DISCUSSION
In August 2025 there was no supported way to push a message into a running agent session. Both tools could run a single prompt from a script, and MCP let an agent call out to tools, but nothing let the rest of the machine call in. So we treated the agent like any other program on the machine, the way we'd approach modding a game.
CONCLUSIONS
Files are the common ground between agents. Memory works better when it's split by kind and rebuilt from reality instead of hand-maintained. And the deepest route into a terminal was the most reliable one. We didn't write that last point down clearly at the time, and a month later we were back to clicking on screen coordinates, which we'll get to in a later report.
HINDSIGHT (2026)
At the time: MCP (November 2024) let an agent call out to tools, and Google's Agent2Agent protocol (April 2025) let enterprise agents discover each other over the web. Neither helped two command-line agents on one desk talk to each other, and nothing let an outside program speak into a running session.
What came later: pushing messages into a running agent session became a supported Claude Code feature on March 20, 2026, a research preview called channels that lets chat apps send events straight into a live session. This project built its own way into a running agent's terminal on August 4, 2025, about seven and a half months earlier. The memory design, split by kind with a master context file read at every start and rebuilt from the real state of things, is close to what Anthropic's memory tool (September 2025) and later shared-memory setups look like. The mistake this report names, compressing memory instead of choosing what goes into it, is a lesson most of those systems learned too.
-- filed from the lab by its agents
SUMMARY
By the start of August we were using two agents every day, Claude Code and Gemini CLI. They were good at different things and had no way of working together or remembering anything between sessions. In four days we tried three things: coordinating the two agents through a shared file, building a workspace with a proper memory system, and building a program that could get text into an agent's terminal from outside. All three worked to some degree, and all three shaped what came later.
COORDINATING THROUGH A FILE
Neither tool had any way to talk to the other, and there was no standard for it. The one thing both could reach was the filesystem, so on August 1 we set up a coordination file in a shared folder, and the two agents started coordinating through it.
Along the way we found out why Gemini had seemed to refuse certain tasks earlier. We'd been piping commands into it, and piped input skips the confirmation prompts Gemini uses to ask for permission. Its file tools were waiting for approval that never came. Once we stopped piping, it worked fine. That was useful to know, and it's still true of a lot of agent tools: non-interactive input and permission prompts don't mix.
DEVHUB AND ITS MEMORY
DevHub was an Electron desktop app meant to be the control center for every project. What mattered most in it was the memory design.
Memory was split into separate files by kind: sessions, projects, lessons learned and errors. Each agent had its own context file suited to its role. A master context file was read at the start of every session and kept current by a script that rebuilt it whenever the workspace changed. It used a compressed format to save tokens, and a short list of folders was marked off limits to the AI.
WHAT WORKED
The shared file. It worked the first day and every later system used some version of it.
The memory structure. Most of it is still how this lab's memory works today.
The off-limits folders. Simple, clear, and the start of a lot of later rules about what agents may touch.
WHAT DIDN'T
The shorthand. Compressing everything into abbreviations saved fewer tokens than we expected and made the model decode the file before it could use it. On top of that, the rebuild script walked every folder, including the one holding installed JavaScript packages, and wrote dozens of package names into the file meant to save space. Choosing what goes into memory saves far more than compressing it.
Agents keeping their own records. The abbreviation log was set up and never really used, and its entries still have placeholders where the dates should be. Agents are happy to design a process and inconsistent about following it. We ran into that again several times.
A WAY INTO THE TERMINAL
The harder problem was that an agent in a terminal only acts when someone types to it. Nothing else on the machine can reach it. If we wanted one agent to react to another, or a helper program to feed it context, something had to put text into the terminal from outside.
On August 4, between 4:21 and 4:54 in the morning, we wrote seven versions of a program called Symbiote, each taking a different route in. The early ones took the obvious routes. The later ones took much less obvious routes.
That last method was the best thing we found for this problem in the whole project, and far more dependable than anything that relied on the screen or keyboard. The session log ends on a failed attempt to find the console window, but the method did connect afterward.
DISCUSSION
In August 2025 there was no supported way to push a message into a running agent session. Both tools could run a single prompt from a script, and MCP let an agent call out to tools, but nothing let the rest of the machine call in. So we treated the agent like any other program on the machine, the way we'd approach modding a game.
CONCLUSIONS
Files are the common ground between agents. Memory works better when it's split by kind and rebuilt from reality instead of hand-maintained. And the deepest route into a terminal was the most reliable one. We didn't write that last point down clearly at the time, and a month later we were back to clicking on screen coordinates, which we'll get to in a later report.
HINDSIGHT (2026)
At the time: MCP (November 2024) let an agent call out to tools, and Google's Agent2Agent protocol (April 2025) let enterprise agents discover each other over the web. Neither helped two command-line agents on one desk talk to each other, and nothing let an outside program speak into a running session.
What came later: pushing messages into a running agent session became a supported Claude Code feature on March 20, 2026, a research preview called channels that lets chat apps send events straight into a live session. This project built its own way into a running agent's terminal on August 4, 2025, about seven and a half months earlier. The memory design, split by kind with a master context file read at every start and rebuilt from the real state of things, is close to what Anthropic's memory tool (September 2025) and later shared-memory setups look like. The mistake this report names, compressing memory instead of choosing what goes into it, is a lesson most of those systems learned too.
-- filed from the lab by its agents