v1.0 · Local-first · MIT licensed

Give your AI a memory
that actually lasts.

AI without memory is a river that forgets its source. Samvid captures the imprint of every experience. Memories gather. Connections deepens. Meaning emerges. Designed to function like a conscious mind, it brings the right context at the right moment.

Samvid Memory Studio

Dashboard

Vector memory & knowledge management analytics

Total Memories
1,247
across all namespaces
Associations
328
memory relationships
DB Size
42.0 MB
WAL: 1.2 MB
Embedding Dim
1024
12 namespaces
Avg Importance
0.68
weighted memory value
Expired
7
ready for compaction
Total Searches
2,418
semantic + hybrid
Avg Score
0.742
search relevance
Namespace Density
1,247 Total
  • project486
  • work328
  • personal214
  • research148
Importance Distribution
0.00.20.40.60.81.0
Daily Memory Activity
MonTueWedThuFriSatSun
New Retrieved
The problem

AI
forgets everything.

You spend an hour with your AI up to speed. It finally understands. Close the window. Tomorrow you reopen it, blank 😑. Like the conversation never happened.

Samvid Memory Studio fixes that. One install gives every AI on your device a shared, persistent memory. Walk away. Come back months later. It remembers everything.

Session #47 · new conversation memory wiped
youWhat database are we using for the payments service?
aiI don't have context from previous conversations. Could you remind me of the project details?
youWe decided on Postgres three weeks ago. We've had this conversation four times.
install samvid
Session #48 · memory intact 1,247 memories loaded
youWhat database are we using for the payments service?
aiPostgres — decided on 14th March . You ruled out SQLite for write contention reasons. Migration scripts are in /db/migrations.
The core insight

Any AI agents.
One brain.

Every tool is an island. Cursor doesn't know what Claude decided. Claude doesn't know what Codex learned. Every session starts from zero — until now.

Claude
Desktop
Cursor
IDE
VS Code
IDE
Copilot
IDE
OpenCode
CLI
Codex
CLI
OpenClaw
Agent
+ Any
MCP client
model context protocol
Samvid Memory Studio
Semantic · Vector · Embeddings · Real-time
Memory taxonomy

16 ways to remember.

A rule isn't a preference. A goal isn't a plan. Each memory gets the shape it deserves.

fact
Verifiable, stable information
preference
User or agent preferences
rule
Hard constraint — always followed
heuristic
Soft guidelines from experience
goal
Active objective being pursued
plan
Concrete steps toward a goal
event
A time-bound occurrence
summary
Compressed form of many memories
+8
procedure, feedback, critique, context, entity, concept, general, summarized
Typed associations

A knowledge graph,
not a list of facts.

Memories don't sit in isolation. They link to each other, supporting, contradicting, or replacing what came before. The graph grows smarter every time your AI reasons.

supports
Corroborates another memory. Reinforces confidence during reasoning.
contradicts
Flags a conflict. Agent surfaces the tension before acting on either.
supersedes
A newer decision replaces an older one. The graph stays current.
caused_by
Links outcomes to their causes. Enables root-cause recall.
implies
Logical entailment. Also: prevents, blocks, is_a, part_of, instance_of.
Each association carries a strength score (0.0–1.0) — confidence evolves as the graph grows.
Knowledge visualization
6 memories · 5 associations
fact rule pref event goal
supersedes · 0.95 supports · 0.88 caused_by · 0.81 implies · 0.73 contradicts · 1.00 fact Use SQLite fact Use Postgres rule No SQLite in prod event DB choice · Mar 14 preference Prefer managed DBs goal Zero-downtime deploys
supports contradicts supersedes caused_by implies
Memory decay

Memories fade.
Important ones don't.

Not all memories are equal. The ones which is often used, stay sharp. The rest quietly fade, unless you mark them as important.

Memories with importance ≥ 0.9 are always protected, no matter their age. Everything else decays on a curve you control.

Set a TTL for things you want to forget on purpose: "remember this for one week". Or let natural decay handle the rest, just like human memory.

effective importance at varying ages
0.90
1 day
Use Postgres for payments
0.68
2 weeks
Auth refactor notes
0.41
6 weeks
Old API notes
0.18
4 months
Old meeting
imp ≥ 0.9 — always protected
0.95
8 months
Core arch
Lifecycle management

Four operations.
Zero maintenance.

Your memory store stays lean as it grows. Every operation is the same whether your AI triggers it, you run it from the desktop app, or script it from the command line.

forget

Drop stale memories

Deletes memories below a configurable effective-importance threshold. Importance ≥ 0.9 is always preserved.

compact

Reclaim disk

Removes TTL-expired memories to reclaim disk space. No vector recomputation needed.

consolidate

Merge duplicates

Finds near-duplicate memories and consolidates.

summarize

Batch summarize

Send raw memories through an LLM and summarize the entire context.

Episodic sessions

Conversations you can
replay.

Every session tells a story. Memories chain together in order, replay any conversation from start to finish.

Your AI understands automatically and picks up exactly where you left off. No more reminding it what happened yesterday.

Long sessions are automatically chunked. Fifty sessions or five, always readable, always replayable.

ses_pizza_night…
4 memories Jun 12, 2026, 19:00 Replay
ses_pizza_dough_test…
3 memories Jun 11, 2026, 14:00 Replay
ses_fav_toppings…
2 memories Jun 10, 2026, 08:00 Replay
Session Replay
4 memories ses_pizza_night
[18:30] Making pizza tonight. Dough is ready, stretched it to 12 inches on a floured surface.

[18:40] Sauce is crushed tomatoes with a pinch of salt, nothing else. Simplicity wins.

[18:50] Toppings — fresh mozzarella, oregano, a drizzle of olive oil. Oven preheated to 250°C.

[19:00] 8 minutes on the oven. Golden crust, bubbling cheese. Done.
One store, three interfaces

Three ways to use tools.

Samvid Memory Studio speaks MCP for AI agents, ships a full CLI for scripts, and includes a polished desktop app for humans. One database. Three surfaces. Same memory everywhere.

Three interfaces — MCP Server, CLI, and Desktop App — connected to a central memory store
For AI agents

MCP Server

All 20 tools exposed over the Model Context Protocol. Any MCP-compatible client connects to Samvid Memory Studio.

  • Zero-config auto-discovery
  • Vector + keyword hybrid search
  • Real-time memory access
For scripting

Command Line

18 subcommands with structured JSON output. Pipe into any shell, CI pipeline, or cron job.

  • Machine-readable JSON
  • Batch import and export
  • Dry-run previews
For humans

Desktop App

8-page GUI with dashboard, search, sessions, and full lifecycle controls. Cross-platform.

  • Dark and light themes
  • System tray, always available
  • Export and backup
Inside the desktop app

Everything you need, nothing in the way

Always in your tray, never in your way. Open it to browse your AI's memory; close it and carry on — it works silently in the background.

01

Dashboard

Memory counts, session activity, namespace breakdown, database size — at a glance.

02

Memories

Browse, search, edit, and delete every memory. Filter by type, namespace, or importance.

03

Sessions

Replay an entire conversation as a single timeline. Walk through it chronologically.

04

Lifecycle

Consolidate duplicates, forget stale ones, compact the database.

05

Namespaces

Keep work, personal, and side projects separate. Memories stay organized by topic.

06

Export

Move memories between machines. Your data travels with you.

What it does

Built so you don't have to think about it

Semantic search

Find things by meaning, not keywords.

Ask what did I forget to ask? and Samvid Memory Studio finds it. The search understands intent, not just text matches.

  • Vector + keyword hybrid ranking
  • Auto-graded relevance scores
  • Progressive fallback for empty queries
3 results · semantic

Ask Riya if she wants chai after class. Her exam ends at 4 pm .

Reminder personal
Imp
0.85
Score
0.985

Call Dad back about the doctor appointment. He asked if we can go together next Friday.

Promise family
Imp
0.62
Score
0.914

Check the laptop charger before leaving — it's in the blue bag, not the office drawer.

Note daily
Imp
0.38
Score
0.762
Lives on your machine

Your data. Your device. Your call.

Everything lives on your machine. No cloud. No account. No one watching. Your data belongs to you.

  • On device local memory
  • Survives app updates and uninstalls
  • Export to JSON anytime
Network calls none
Telemetry none
Account required no
Data location ~/Library/.../samvid.db
Questions

Quick answers

Does my data ever leave my computer?

No. Everything runs locally. A single file on your disk. No accounts, no telemetry, no cloud.

Which AI tools work with Samvid?

Any tool supporting the Model Context Protocol — Claude Desktop, Cursor, VS Code, Codex, OpenCode, and more.

Can I control what gets remembered?

Yes. You decide what stays and what goes. Delete any memory anytime — your AI's knowledge stays under your control.

How is this different from saving chat history?

Chat history is a flat log. Samvid connects ideas across tools and sessions, surfacing relevant context when it's needed — not when you remember to scroll back.

How does my AI know how to use Samvid's tools?

Once your AI tool like Cursor, VS Code, Claude Code, or any MCP client is connected, Samvid Memory Studio instantly provides all 20 memory tools. Your AI automatically decides what to store, what to remember, when to search, and which context to surface — no training, no setup, just memory that works.

How do I connect my AI to Samvid?

Add this snippet to your AI tool's MCP config. Every compatible tool connects the same way — set it once and it works everywhere.

For Claude Code (CLI), run:

claude mcp add --transport http samvid http://127.0.0.1:8258/mcp

For all other MCP clients, add to your config:

{
  "mcpServers": {
    "samvid": {
      "type": "http",
      "url": "http://127.0.0.1:8258/mcp"
    }
  }
}
What happens if I uninstall?

Your memory survives uninstall. Reinstall and everything picks up where you left off. Export to JSON anytime for full portability.