Abstract
Agents working over professional document collections face a structural failure: load entire files into context (prohibitively expensive) or search by filename alone (unreliably wrong). The Hippo Campus Protocol (HCP) separates memory from reasoning. Files are prepared in a local searchable library ahead of time; when a question is asked, agents receive compact triage cards — summaries, anchors, confidence signals, location references — before committing to full reads.
The agent’s document problem
When a user asks “find the plumber invoice from March” or “what does the contract say about termination?”, an agent without a memory layer has four bad options:
- Read everything — hundreds of thousands of tokens.
- Filename search — low cost, high failure rate.
- Query-time LLM — re-infer on every question; slow and expensive.
- Thin file wrapper — returns paths; agent still reads whole files.
The agent needs a decision layer before a reading layer. HCP formalises that split.
Design thesis
Memory is not reasoning. Prepare the files; return the source; hand only the relevant context to the model you already use.
Why “Hippo Campus”
The hippocampus encodes experiences, binds content to context, and retrieves from incomplete cues — it does not “think”; it makes thinking possible by supplying the right trace at the right time. A campus is an organised memory space: buildings (files), maps (chunks), directories (types), transit rules (scope, permissions, confidence). Agents navigate by cue, not by memorising every room.
What agents receive
When a question is asked, HCP returns short result cards: the most relevant passages, which file they came from, their page, cell, or timecode, and how confident the match is. Each result is designed to fit in a few hundred tokens, not tens of thousands.
When a short result is not enough, the agent can request more from the source. The default is a small, linked result; loading more is an explicit choice, not an automatic whole-file dump.
Local connection for AI tools
IDE agents and desktop assistants connect to HCP on the customer’s machine or private network. The searchable library is not sent through a Find and Seek cloud API. You keep your AI tool; Find and Seek supplies its memory.
Campus isolation
Each department, matter, or client can have its own search boundary. Permissions are checked before anything reaches an agent. Our multi-team benchmarks test this pattern across 1, 3, and 6 permission groups.
Limitations (honest)
Short result cards are the default, not a ceiling. Agents that need more can request it. A limited result may miss a detail that only a full-file read would catch — we document those limits in measured benchmarks rather than hiding them.
Savings apply when agents use Find and Seek to answer questions about connected files — not for general chat or queries that bypass it.
Find and Seek
Find and Seek implements the HCP memory plane: organise files ahead of time, return the source, and run approved analysis locally. Measured outcomes are on the Evidence page. Product overview on Product.