GPS gives coding agents a memory that actually works

Every developer who uses a coding agent knows the loop. You correct a mistake. The agent apologizes. Next session, same mistake. You explain the rule again. The agent stores it in a bloated CLAUDE.md that grows until the model ignores it. Repeat.

GPS, an open-source tool launched this week by Invariance AI and listed on Product Hunt, tries to break that loop. Instead of dumping instructions into a flat file, GPS anchors memories to the specific files, functions, and symbols they describe. When an agent prepares to edit a module, GPS surfaces only the memories tied to that module. Nothing else.

The problem GPS addresses is structural. Current coding agents from Anthropic (Claude Code), OpenAI (Codex), and Cursor treat each session as a fresh context window. They can read a repo, but they cannot remember what they learned last time. The standard workaround is the CLAUDE.md file, a project-level prompt that developers fill with rules, gotchas, and conventions. As the file grows, the model’s attention dilutes. A rule about logging PII in the payment module sits next to a rule about test command syntax. The agent cannot tell which rule applies when.

Hardik Singh, the Invariance AI engineer who built GPS, described the problem in the Product Hunt launch thread. “I kept noticing my agents making the same mistakes across sessions,” Singh wrote. “I’d correct something, explain a rule, show it where the tests were, and the next run it was gone. Back to square one. Every session I was burning tokens just re-explaining context that should’ve already been there.”

GPS stores memories as structured records tagged to file paths and symbol names. A memory that says “never log user IDs in this module” is attached to that module’s path and the relevant function signatures. When the agent is about to edit that code, GPS retrieves exactly those memories. It does not perform keyword search. It does not dump the entire memory store into the system prompt. It uses the agent’s planned edit target to select what is relevant.

That design choice matters for token economics. A typical project memory file might run 2,000 to 5,000 tokens. A GPS retrieval against a single file might return 200 to 500 tokens. For a team running dozens of agent sessions per day, the savings compound. More importantly, the signal-to-noise ratio improves. The agent receives dense, relevant context instead of a document it mostly ignores.

The tool also lets agents file their own memories mid-task. If an agent discovers an edge case, fails on a command, or learns something worth remembering, it saves the memory automatically. The next session starts smarter than the last one ended. That is a subtle but important shift. Most memory systems treat the developer as the sole source of truth. GPS treats the agent as a contributor to institutional knowledge.

Binu George, a commenter on the Product Hunt thread, raised a sharp question about GPS’s scope. “The harder case is when a rule is cross-cutting,” George wrote. “Say, a constraint that applies to any function that touches a particular data model regardless of where it lives. Does the memory layer handle that kind of relational mapping, or is the retrieval currently path/symbol-bound?” Singh’s answer was not included in the extracted source, but the question points to a real limitation. Many coding standards cut across file boundaries. A rule about data validation applies everywhere a certain schema appears. GPS’s current architecture, which anchors memories to specific files and symbols, may struggle with those patterns.

That limitation does not diminish what GPS gets right. The tool solves a specific, painful problem: the amnesia that plagues every session-based coding agent. It does so with a clean architecture that is local-first, CLI-first, and built for the major agent platforms including Claude Code, Codex, Cursor, and any tool supporting the Model Context Protocol (MCP).

The broader implication for the AI engineering market is this. As coding agents move from novelty to daily tool, the quality of their memory infrastructure will determine whether they save time or waste it. A coding agent that forgets everything between sessions is a coding agent that requires constant supervision. A coding agent that learns, remembers, and recalls relevant context is a coding agent that can work autonomously.

GPS is not the only tool pursuing this vision. Anthropic has experimented with persistent memory features in Claude Code. Several startups are building agent memory layers as standalone products. But GPS’s open-source approach and its focus on symbol-anchored retrieval give it a distinctive position. It is free, transparent, and designed to integrate rather than replace.

The question for teams adopting GPS is whether the cross-cutting memory problem proves tractable. Singh and Invariance AI will need to show that GPS can handle rules that span modules, not just rules that live in one file. If they can, the tool becomes a standard piece of the AI coding stack. If they cannot, it remains a useful but partial fix.

For now, GPS is a reminder that the most valuable AI tools are often the ones that solve boring problems well. Memory is not glamorous. But every developer who has corrected the same agent mistake twice knows exactly how much it costs.