Scott Taylor
Co-Founder & CEO
Profile summary: Memco • Co-Founder & CEO
Memco
Memco provides a governed memory layer for AI agents, converting successful fixes, failed paths, and human corrections into reusable lessons that persist across models, tools, and teams. The platform captures agent traces, routes them through a human review gate, and makes scoped lessons available to future sessions, eliminating the cold-start tax where agents repeatedly rediscover the same knowledge. In independent benchmarks, agents using Memco achieved 2.6× the task success of static-RAG retrieval without any weight updates or fine-tuning.
- Artificial Intelligence
- AI Agents
- Developer Tools
- Software Only
Updated 2 months ago
Funding
Funding not disclosed
across 1 round
RoundDateAmountValuationInvestors
Founders
Product
Problem
AI agents repeat work when lessons stay trapped in a single session. Each new run rediscovers repository quirks, failed paths, and human corrections, forcing companies to pay a "cold-start tax" in tokens and senior engineering time. Context windows reset when sessions end, so knowledge that could compound across runs is lost unless it is deliberately captured and governed.
Solution
Memco provides a governed memory layer that turns validated agent work into reusable, organization-owned lessons. The platform captures a trace from each agent run—including fixes, dead ends, and human corrections—then routes it through a review process where teams promote useful parts into scoped, owned lessons. Before the next agent starts spending tokens, it recalls relevant lessons from shared memory, allowing it to begin ahead of the first run. Memco is model-agnostic, IDE-agnostic, and harness-agnostic, working across Claude, OpenAI, Gemini, Llama, and other stacks, with support for Cursor, Windsurf, Zed, JetBrains, VS Code, and multiple agent harnesses. The system includes provenance on every lesson, role-based access control, and retirement as a first-class action, ensuring memory stays accurate and within organizational boundaries.
Target Audience
Primary customers are engineering teams running coding agents on shared repositories, along with private equity operating partners, enterprise CIOs/CTOs, and customer support leaders who need validated agent work to persist as governed organizational memory across teams and tools.
Features
- Trace-to-lesson pipeline that captures agent runs, promotes useful parts through human review, and stores scoped lessons with provenance (run, reviewer, scope)
- Model-agnostic architecture supporting Claude, OpenAI, Gemini, Llama, DeepSeek, and Qwen, with compatibility across major IDEs and agent harnesses
- Scope boundaries for personal, team, and organization-level memory with RBAC for read, write, approve, and retire permissions
- Retirement as a first-class action, allowing stale or outdated lessons to be explicitly retired rather than lingering in retrieval
- Self-host or VPC deployment options for enterprises requiring data residency and security boundaries
- Published benchmark results showing 2.6× task success versus static-RAG baselines and 88% retention of learned solutions, with zero weight updates or fine-tuning