A research base that compounds

> Scattered findings into a graph your AI walks.

type: playbookfor: [researchers]

Research is allergic to chat windows. Build it somewhere it can grow. We use Basic Memory for our own competitive intelligence and product research: each source becomes a structured note, each finding becomes a tagged observation, and typed relations link papers and companies into a graph. Months later, semantic search finds the insight you forgot you'd written.

Read the research guide › Start free ›

## The problem

Your research scatters and disappears.

Research insights end up scattered across chats, PDFs, and a Notion page you opened twice. You re-read the same papers, lose the thread between sessions, and can't see how findings connect. When the question comes back six months later, the work is gone with the chat history.

## How it works in Basic Memory

The workflow.

Capture each source as its own note. Frontmatter for type and tags, prose sections for analysis, tagged observations like [strength], [weakness], [opportunity], [insight] for the structured findings.
Add typed relations as you go. `competes_with`, `similar_to`, `integrates_with`, `synthesizes`. Relations make the graph queryable, not just searchable.
When you need to pull findings together, write a synthesis note: a summary that draws on several sources and keeps a typed link back to each one (`synthesizes [[Source A]]` / `synthesizes [[Source B]]`). It stays wired to every source it draws from.
Six months later: semantic search finds the right note, build_context walks the synthesizes edges back to the underlying sources, and your AI helps you write the new piece on top of the old graph instead of starting cold.

## What you get

The outcome.

Your research compounds instead of scattering. Every source and every synthesis joins one connected base, so a finding from six months ago is one search away from today's question. The notes stay plain Markdown you own, readable by you in any editor and usable by any AI tool you connect.

## In practice

How we run competitive intelligence on the AI memory space.

Our own competitive intelligence lives in a Basic Memory project. Each competitor gets a deep-dive note with structured sections and tagged observations. A summary note synthesizes across the deep dives via typed relations. Six months later, when the market shifts and a partner asks about it, we walk the graph instead of re-reading everything.

research/competitive-intel/Deep Dive - Mem0.md (source note, excerpt)

Each competitor is a structured note. Frontmatter for type and tags, prose sections for company overview and architecture, and tagged observations at the bottom that capture the findings in queryable form. Categories like [strength], [weakness], [opportunity], [watch] aren't decoration; they make it possible to ask the workspace later: 'show me every [opportunity] across every deep dive.'

---
title: Deep Dive - Mem0
type: note
tags:
- competitor
- mem0
- agent-memory
- b2b
- infrastructure
---

# Deep Dive: Mem0

## Company Overview

**Founded:** Late 2023
**Funding:** $24M total (Seed + Series A, Oct 2025)
**Team Size:** 4 people (at funding)
**HQ:** San Francisco, CA

## Technical Architecture

Mem0 uses a three-store hybrid architecture:
1. Vector Database — embeddings for semantic similarity
2. Key-Value Store — structured data and metadata
3. Graph Database — relationships between entities

## Competitive Strengths

- [strength] Distribution — AWS exclusive partnership is massive
- [strength] Open source momentum — 41K stars, massive adoption
- [strength] Model agnostic — Works with any LLM

## Competitive Weaknesses

- [weakness] Small team (4 people) limits execution capacity
- [weakness] Cloud-first approach — privacy-conscious customers
  may hesitate
- [weakness] No local-first option like Basic Memory

## Observations

- [threat] Mem0 is the most well-funded, highest-traction direct
  competitor
- [insight] Their "memory passport" vision aligns with multi-agent
  future
- [opportunity] Local-first + privacy narrative is underserved by
  Mem0
- [watch] AWS partnership could lock in enterprise segment

## Relations

- competes_with [[Basic Memory]]
- similar_to [[Zep AI]]
- integrates_with [[CrewAI]]
- integrates_with [[LangChain]]

research/competitive-intel/Competitive Intelligence Summary - February 2026.md (synthesis, excerpt)

The summary synthesizes across the deep dives via typed `synthesizes [[Deep Dive - X]]` relations, leaving each one intact. The exec summary, the funding-landscape table, and the competitive matrix all live as observations. A future agent following the relations from this note walks straight back to the underlying source notes.

---
title: Competitive Intelligence Summary - February 2026
type: note
tags:
- strategy
- competitive-intelligence
- market-analysis
- summary
---

# Competitive Intelligence Summary - February 2026

## Executive Summary

1. Mem0 is the dominant direct competitor — $24M funding, AWS
   partnership, 41K GitHub stars
2. Zep differentiates on temporal graphs — SOC 2/HIPAA compliance,
   enterprise focus
3. Letta owns academic credibility — Berkeley pedigree, self-editing
   memory research
4. B2C personal AI is struggling — Limitless acquired by Meta,
   consumer hardware failing

## Basic Memory's Moat

Differentiators no competitor has:
1. Local-first + MCP — true privacy with Claude integration
2. Obsidian ecosystem — PKM user base as distribution
3. Human-readable — Markdown, not databases
4. Developer-owned — your notes, your format

## Observations

- [strategy] Local-first + MCP + Obsidian is unique positioning
- [opportunity] Mem0's cloud-first approach leaves privacy gap
- [opportunity] No competitor has PKM integration
- [insight] B2C AI hardware is failing; B2B developer tools more
  defensible
- [watch] Enterprise compliance (SOC 2, HIPAA) becoming table stakes

## Relations

- synthesizes [[Deep Dive - Mem0]]
- synthesizes [[Deep Dive - Zep AI]]
- synthesizes [[Deep Dive - Letta (MemGPT)]]
- synthesizes [[Deep Dive - Rewind AI Limitless]]
- synthesizes [[Deep Dive - LangChain Deep Agents]]
- informs [[SPEC-38 Basic Memory Backends]]

What the graph does six months later

Months after the original research, a question comes in: 'Is local-first still a defensible angle, or did the market catch up?' Instead of re-reading every deep dive, the agent searches the graph and walks the relations to assemble the answer from notes already written. The synthesis is a node, not a dead end.

# Agent flow on a new question

> Is local-first still a defensible angle for AI memory?

search_notes(query="local-first privacy", tags=["opportunity"])
  → 3 hits across Deep Dive - Mem0, Deep Dive - Zep AI,
    Competitive Intelligence Summary

build_context(url="memory://research/competitive-intel/
  competitive-intelligence-summary-february-2026")
  → walks synthesizes edges → loads the 5 deep dives plus
    the moat section from the summary

Agent answers from the existing graph, citing the source notes
by permalink. No fresh research run; six months of accumulated
analysis is already there.

Real synthesis

The moat, as written in our own research workspace

“Local-first + MCP + Obsidian is unique positioning. No competitor holds all three. The cloud-first players (Mem0, Zep, Letta) give up readability and ownership; the plain-files players (a folder of Markdown, AGENTS.md, OKF bundles) give up the index that makes the files queryable. We hold both ends.”

Synthesized in our own competitive-intelligence summary, which `synthesizes` five deep-dive notes via typed relations. The synthesis above is a node in the graph, with the deep dives that informed it one hop away.

## FAQ

Common questions.

How do I use AI to build a research knowledge base?

Capture each source as its own note with frontmatter and tagged observations (like [strength], [insight], [opportunity]). Add typed relations (`competes_with`, `synthesizes`, `similar_to`) as connections emerge. Your AI reads and writes the same files over MCP, so it helps structure the notes, finds related work via semantic search, and walks the graph when synthesizing.

How is this different from saving sources to Notion or a Zotero library?

Two things. Tagged observations and typed relations make the structure queryable: an agent can search for all [opportunity] observations across the workspace, or follow `synthesizes` edges from a summary back to its sources. And the notes are plain Markdown you own, usable by every other MCP-connected agent and editable by hand in any editor.

If I summarize across sources, does it stay linked to them?

Yes. A synthesis note keeps a typed link back to every source it draws from (`synthesizes [[Source]]`). Months later, opening it lets you or your AI jump straight back to each one, so you never lose track of where a conclusion came from.

Is my research data portable?

Every note is plain Markdown you own. Readable in any editor, versionable in Git, exportable at any time. The graph is just typed wiki-links between files, so the same links resolve in any editor that reads Markdown.

Try this playbook.

Free and open source to run locally. Two minutes to connect your first AI tool.

Read the research guide › ← All playbooks