HiveMem
Your second brain — and it stays yours. Forever. Local.
A sovereign personal knowledge system. The conversations, decisions, documents, and half-formed thoughts you produce across Claude, ChatGPT, Gemini, Copilot — and the files you accumulate in real life — all come home to one place that outlives any vendor and obeys only you.
Why HiveMem exists
When you think hard today, you often think with an LLM in the loop. School, work, authorities, court cases, taxes, family, health, relationships — these conversations contain your most private thinking. More intimate than any diary.
And then they evaporate:
- Your subscription lapses or you switch providers → history gone
- The provider retires a model or rewrites their ToS → answers no longer reproducible
- An account ban, a provider going under, a country blocking the service → everything lost
- The data sits on a vendor's servers, fed into training, served on subpoena, exposed in the next breach
HiveMem is built around the opposite stance:
- Sovereignty — Your data lives in your instance. Postgres + SeaweedFS, on hardware you control. No vendor sees the contents unless you explicitly route a single LLM call through them.
- Persistence — Everything is append-only with
valid_from/valid_until. No subscription change can revoke access. No retention policy you didn't author can delete what's yours. - Portability — A HiveMem instance packs into one encrypted archive (Postgres dump + binary store + config) and restores anywhere. Vendor lock-in: zero.
- Aggregation — What you write in Claude.ai, ChatGPT, Gemini, Claude Code, Copilot lands in HiveMem too. Those tools become front-ends; HiveMem holds the truth.
- Privacy by realm — Strict separation per life area
(
legal,medical,private,work). Per-realm routing rules: anything touching authorities or health stays on local models, never reaches a cloud provider.
Knowledge doesn't rot here
The long-term goal is a periodic agent — the Queen — that wakes on a schedule, surveys your knowledge, and dispatches specialized worker agents (Bees) to flag isolated cells, stale facts, duplicate candidates, and realms drifting from their blueprint. Everything risky stays a proposal that flows through the existing approval workflow; you keep the kill switch.
Today the Queen and the isolated-cell Bee already run on the Vistierie agent
runtime — scheduled (cron), dispatched with per-run cost accounting and a
per-tenant kill switch — and their proposals flow through the approval workflow
as pending tunnels. The same runtime also hosts the inbox archivist,
which fully-automatically files cells out of the inbox staging realm (triggered
right after enrichment settles, plus a daily safety-net run) — its moves are
logged with a reason and reversible, not proposals awaiting approval. An
admin-only Queen log UI (/queen) shows run history, event timelines, and
the proposal queue for all three agents. Still to come: a conversation UI that
teaches the Queen your per-realm preferences, and further Bee types (stale-fact,
duplicate-cell, blueprint-drift).
→ Roadmap — what's planned, what's partial, and the order of work.
→ Scientific foundations — the cognitive-science and PKM theory HiveMem's design is built on (Working Memory, Cognitive Load, Extended Mind, Forgetting Curve, Zettelkasten, PARA).
Docker images: ghcr.io/visterion/hivemem:main for the rolling main branch, plus semver tags such as ghcr.io/visterion/hivemem:9.1.5 for cut releases.
Highlights
- 6-Signal Ranked Search — Semantic similarity, keyword, recency, importance, popularity, and graph proximity — combined into one ranked result.
- Temporal Knowledge Graph — Facts with
valid_from/valid_untiland multi-hop graph traversal; query the graph as it stood at any point in time. Subject canonicalization (kg_alias) collapses fragmented subject strings ("Acme Inc." vs "Acme") into one canonical entity, retro-migrates existing facts, and is used bykg_add/entity_overviewto resolve subjects on both the write and read path. - Progressive Summarization — Four layers per cell: content, summary, key points, and insight. Never lose nuance.
- Document & Scan Pipeline — the end-to-end picture of how any file becomes searchable knowledge: two entry points (watched folder + REST upload), one shared ingest core (hash → parse → dedup → store → cell), and four async enrichment paths (OCR · Vision · Kroki · Summarizer). The map that ties the doc features below together.
- Long cells stay searchable — auto-summarizer turns multi-page documents into curated summaries that are embedded for semantic search; cost-capped, opt-in.
- Scanned PDFs become searchable — Tesseract OCR extracts text from scan-only PDFs; combined with the auto-summarizer, even paper-mailed documents are findable by semantic search.
- Consumption folder — auto document separation — Drop a stack of mixed scans into a network folder (SMB); HiveMem OCRs each page and uses a Vistierie LLM agent to split multi-page batches into individual documents by content — no separator or barcode sheets. The USP over Paperless-ngx; live in production.
- Document-Type Extraction — invoices, contracts, and other typed documents are auto-classified during summarization; typed facts (vendor, amount, parties, dates) land in the knowledge graph. The summarizer also detects content language and (auto-applying a / tag) and parses each document's own date into the cell's , so timeline and sort-by-date views reflect the document date, not the ingest time.
Feature Status
Honest snapshot of what is shipping today versus what the surrounding prose describes as the long-term shape. See the roadmap for details on every 🟡 / 🔴 row.
Documentation
License
HiveMem is fair-code licensed under the Sustainable Use License. Free for personal and internal business use. See LICENSING.md for details.