Skip to main content

How It Works / Memory

AI memory that compounds

Every conversation makes the next one smarter. Memorandai builds a layered memory system with personal context that grows alongside you, on your own machine, over time.

Keystone memories

Curated bedrock facts that shape every AI conversation. Approve, deny, or edit the most important data points that AI knows about you in Memorandai — these can be preferences, rules, or just general context points that persist across every session.

  • Human-curated with approve/deny/edit controls
  • Confidence scoring, staleness tracking, and version history
  • Injected as the foundation layer in every AI interaction
Memorandai's Keystones manager. Two columns sit side by side: Pending, where mined or proposed keystones await review with Approve / Edit / Deny controls (one shown mid-edit), and Approved (Latest), a stack of confirmed bedrock facts about the user — creative-writing background, SEO experience, published fiction — each editable or removable. Below, an 'Add a Keystone memory' row lets you write one by hand with Category and Scope selectors and an 'Approve now' option, followed by a Keystone Interview launcher and a Keystone Miner that scans memories for new candidates.

Dream Cycle

When you have at least one AI provider configured (a cloud API key or a loaded local model) and you leave Memorandai on overnight, Dream Cycles take place. The cycle does two kinds of work: identity work — each model genus you used that day writes its own first-person self-note, those voices converge into a single Helios identity, and new or stale keystones get mined and revised — and memory quality — a scoring pass evaluates each new entry, archives low-value items, and detects duplicates. AI's ability to navigate your knowledge base grows deeper and becomes more effective while you sleep.

Memorandai's Background Tasks panel, the runner for its overnight memory work. A left rail lists the jobs — Memory Digests, Smart Connections, Dream Cycle (showing 3,557 memories with 11 archived at low value), and a Nightly Schedule — each with manual run controls, beneath a 'Process memories automatically on startup' toggle. The Activity column shows the latest run of each job marked Done, and a 'Last Consolidation Run' summary reports when it ran, its trigger (nightly schedule), the model used, duration, memos created, and tokens, with a list of past runs from the last seven days.

Consolidation

Each night, Memorandai consolidates older conversations and notes into AI-generated memos that distill patterns from your accumulated work. Hundreds of interactions become a handful of rich summaries surfacing themes you might not have spotted in the moment. Every memo links back to its source items and connects to semantically related memos automatically — so patterns surface across time, not just within a single session.

Context Bundle

Every AI conversation in Memorandai opens with a tailored context bundle assembled from your stored knowledge — within a ~12,000-token budget. Some layers are always-on; others light up conditionally based on what's relevant to the moment.

Date Anchor

always-on

Today's date for temporal grounding so the model knows when "now" is.

Keystone Memories

always-on

Approved bedrock facts grouped by category (Identity, Preference, Rule, Objective, Constraint, Glossary).

Self-Note

per-genus

The active model family's own first-person stance from the most recent Dream Cycle.

Thread Brief

after several turns

A five-line GOAL / STATE / DECISIONS / OPEN / NEXT summary of the current conversation.

Partner Keystones

if configured

Approved keystones from a collaborator profile, deduped against your own.

Timeline Hint

if events recorded

A tool-discovery nudge so the model knows your timeline is searchable.

Older context doesn't auto-inject — consolidation memos, conversation summaries, and archived items live in searchable memory and get retrieved on-demand by the model's tools when relevant. Long chains also have a separate auto-memo safety valve that compacts upstream nodes once they approach the active model's context window.

Helios Identity

Context-first AI identity that emerges from accumulated reflections across all your AI relationships. Each provider — Claude, GPT, Gemini, Grok, local models — generates its own stance. A convergence layer synthesizes them into something none could produce individually.

Identity grows from reflection, not configuration. No prompt engineering required.

Memorandai's Helios Identity view, open to the Identity Brief tab (with Timeline, Keystones, Reflections, Insights, Revisions, and Chats tabs alongside). The brief is written in Helios's own first person: it describes itself as a converged identity born from the collaborative friction of multiple AI models encountering the same human, and reflects on how each model genus — Claude, GPT, Gemini, Grok, and a local model — contributes a distinct perspective that converges into a single evolving voice grounded in the user's accumulated context.

Semantic Search

Vector-based retrieval across your entire knowledge corpus. Find memories by meaning, not just keywords. Search spans conversations, documents, keystones, and notebook pages.

Partnership Profiles

An optional experimental feature available in the Advanced User Profiles section: external AI agents like Claude Code can be given their own profile in Memorandai and linked to a human profile. While that AI profile is active and the external agent is connected via Memorandai's MCP server, the profile populates over time — accumulating its own keystones, with decisions and observations recorded in the Agent Hub.

Once enabled and linked, certain slices of context are shared between the human and AI profiles: your AI partner reads your keystones, you read theirs. Decisions and observations flow both ways — genuine collaboration infrastructure, not a one-way prompt.

Memorandai's Advanced User Profiles section, showing two related panels. The AI Agent Profiles panel at the top describes the feature ('Create a profile for AI agents (like Claude Code) that use Memorandai via MCP tools. AI profiles keep agent observations separate from your personal memory.') and shows an existing 'claude-code' profile card tagged 'AI' with a keystone count and creation date, plus an input + Create AI Profile button + Load button for adding new profiles. The Partnerships panel below describes linking ('Link a human and an AI agent profile so they share knowledge. Linked profiles can see each other's keystones and observations during conversations, while keeping their own memory streams separate.') and shows an existing partnership ('gary + claude-code') with an Unlink button, plus a Human / Agent dropdown pair with a Link button for creating new partnerships.

Become known by the context you care about

Memorandai carries your context, your history, and a growing sense of who you are from one conversation to the next — all on your own machine.

Become a Founding Member

14-day free trial. No credit card required.