I accidentally hit SOTA on agentic memory by using AI companions

Developer tools startup graphCTX claims its local memory system for AI coding agents achieves state-of-the-art performance, retrieving repo facts in ~1ms with flat latency even at 5,000 facts, outperforming cloud-based alternatives like Supermemory. The tool captures coding facts from repos, binds them to git state, and promotes only durable knowledge, aiming to reduce context repetition for developers.

Built to be the best context layer for AI coding agents. graphCTX keeps repo knowledge close to the work: commands, conventions, decisions, and hard-won fixes. Developers spend less time re-explaining context and more time shipping, with local memory that is fast, private, and reliable. curl -fsSL https://graph.coder.company/install | sh See the benchmarks → bench user: what's the deploy command? agent: I don't see one in the repo — you'll need to tell me. user: what's the deploy command? agent: ./scripts/ship.sh --canary --wait mem:9f3a2c Same agent, same prompt. graphCTX gives it the repo memory developers otherwise repeat. How graphCTX manages memory. The system is deliberately narrow: capture reliable coding facts, keep them valid as the repo changes, and return only the context the agent can use. Extract trusted coding facts graphCTX reads package scripts, lockfiles, CI, editorconfig, AGENTS.md, and session episodes so the memory base starts from repo evidence, not model guesses. Bind memory to git state Facts are valid against commits and branches. When code changes, memory moves with the DAG instead of drifting on wall-clock timestamps. Promote only durable knowledge Session details can become workspace or user memory only after evidence gates. Secrets and low-trust prose stay out of durable context. Choose the smallest useful set A relevance gate scores topic drift, entities, and file scope so the agent gets specific context instead of a noisy memory dump. Attach provenance to every recall Returned memory is compact, budgeted, and tagged with source provenance, making it easier for developers to trust and audit what the agent uses. Benchmarked against Supermemory. Same coding-fact set, same queries. graphCTX runs locally and answers in ~1ms; a live Supermemory search round-trip measured ~494ms p50 , so recall stays fast and predictable during development. Reproducible: graphctx compare --deep . Latency stays flat at scale Per-prompt retrieval p50/p95 as the workspace grows. Indexed lookup plus a bounded semantic re-rank keeps the hot path at ~1ms — a 5,000-fact monorepo retrieves as fast as an empty one. Reliable after compaction Post-compaction solve rate across 14 coding tasks. graphCTX restored the needed repo fact in every run. // graphctx eval run --arms A,B,C // scope: this compares local latency + cost on direct coding-fact retrieval for developer workflows. Supermemory targets general/conversational memory with cloud connectors, cross-document reasoning, and neural embeddings that graphCTX doesn't attempt. Start using repo memory in 30 seconds No account. No API key. No cloud setup. Install the CLI, connect your agent, and give every session repo-aware memory. curl -fsSL https://graph.coder.company/install | sh Prefer npm? npm i -g graphctx $ curl -fsSL https://graph.coder.company/install | sh install the CLI detects Node / Bun $ graphctx install claude wire your agent claude · cursor · opencode · generic $ graphctx doctor verify graphCTX is connected