Skip to content

ValorBrain or Letta?

What each one delivers, row by row. Every claim about Letta comes from its own documentation, with a link and a date; every claim about ValorBrain comes from our public pages.

Where ValorBrain stands out

  • MCP with OAuth and dynamic client registration

    ValorBrain: yes · Letta: no (Bearer API key only; no OAuth yet)

  • Hooks in coding agents (Claude Code, Codex, Cursor…)

    ValorBrain: yes · Letta: partial (Claude Code plugin (demo, not for production))

  • Source connectors (Notion, Drive, Slack, GitHub…)

    ValorBrain: yes · Letta: partial (GitHub and Slack for cloud agents; no ingestion)

  • Hybrid search (keyword + vector)

    ValorBrain: yes · Letta: partial (Hybrid only for messages (Cloud) or via mod)

  • New information supersedes the old

    ValorBrain: yes · Letta: partial (Agent/dreaming rewrites memory; no formal invalidation)

  • Read permission per source or document

    ValorBrain: yes · Letta: partial (Per agent/conversation, not per source)

Where Letta stands out

  • Open source (OSI license)

    ValorBrain: no · Letta: yes (Apache-2.0 (Letta harness / letta-code))

  • Paid plan purchasable online today

    ValorBrain: no · Letta: yes (Buy credits and upgrade online)

ValorBrain documents it, Letta doesn't

  • Declared data region

    ValorBrain: yes · Letta: not documented

  • Knowledge graph or time-aware facts

    ValorBrain: yes · Letta: not documented (not documented (Markdown+git memory))

  • Audit log

    ValorBrain: yes · Letta: not documented (not documented (git history of memory only))

Both have it

  • Managed cloud

    ValorBrain: yes · Letta: yes (Letta Cloud (chat.letta.com, API))

  • Self-host, on-prem or VPC

    ValorBrain: yes · Letta: yes (OSS self-host (local runtime / App Server))

  • Hosted MCP server (HTTP)

    ValorBrain: yes · Letta: yes (https://api.letta.com/mcp (HTTP, bearer))

  • Team: members, roles, invites

    ValorBrain: yes · Letta: yes (Organization, Admin/Editor/Analyst roles)

  • Free plan

    ValorBrain: yes · Letta: yes (USD 0/mo; 3 agents, BYOK)

  • Published price (model and entry)

    ValorBrain: yes · Letta: yes (subscription+usage; Pro USD 20/mo)

  • Vendor-published benchmark

    ValorBrain: yes · Letta: yes (LoCoMo 74.0% (GPT-4o mini, Filesystem))

How ValorBrain describes itself

Shared memory that keeps work moving when the person, the agent or the model changes. source↗

How Letta describes itself

AI research lab building stateful agents that remember everything and learn continuously. source↗

Full table

Row by row, with the source of every cell.

yes = documented and available today · partial = exists with a limit (see detail) · no = the source shows it is absent · not documented = the vendor doesn't say · ↗ = the cell's source

Deployment and compliance

Deployment and compliance
CriterionValorBrainLetta
Managed cloudyessource↗SaaS at valorbrain.valor.digitalyessource↗Letta Cloud (chat.letta.com, API)
Open source (OSI license) to weighnosource↗Closed source, for securityyessource↗Apache-2.0 (Letta harness / letta-code)
Self-host, on-prem or VPCyessource↗On-prem or dedicated instance (Enterprise, under contract)yessource↗OSS self-host (local runtime / App Server)
SOC 2 reportnosource↗In preparation; LGPD-aligned, DPA publishednot documentedsource↗Trust center lists no certification or report
Declared data regionyessource↗Documents and vectors in Brazilnot documentedsource↗

How the agent connects

How the agent connects
CriterionValorBrainLetta
Hosted MCP server (HTTP)yessource↗mcpbrain.valor.digital/mcpyessource↗https://api.letta.com/mcp (HTTP, bearer)
MCP with OAuth and dynamic client registrationyessource↗OAuth 2.1 with registration_endpoint (DCR)nosource↗Bearer API key only; no OAuth yet
Hooks in coding agents (Claude Code, Codex, Cursor…)yessource↗Claude Code, Codex, Gemini CLI, Kiro, Cursorpartialsource↗Claude Code plugin (demo, not for production)
Source connectors (Notion, Drive, Slack, GitHub…)yessource↗Notion, Drive, Slack, GitHub, Linear and more; paid planspartialsource↗GitHub and Slack for cloud agents; no ingestion

Retrieval and updates

Retrieval and updates
CriterionValorBrainLetta
Hybrid search (keyword + vector)yessource↗BM25 + vector + RRF + cross-encoderpartialsource↗Hybrid only for messages (Cloud) or via mod
Knowledge graph or time-aware factsyessource↗Keyed facts with a date (as_of), queryable over timenot documentedsource↗not documented (Markdown+git memory)
New information supersedes the oldyessource↗Keyed fact wins over old prosepartialsource↗Agent/dreaming rewrites memory; no formal invalidation

Team and governance

Team and governance
CriterionValorBrainLetta
Team: members, roles, invitesyessource↗Plans by number of people; agents don't countyessource↗Organization, Admin/Editor/Analyst roles
Read permission per source or documentyessource↗Per-source ACL with validity and reason; enabled per tenantpartialsource↗Per agent/conversation, not per source
Audit logyessource↗Immutable mutation trail + read lognot documentedsource↗not documented (git history of memory only)

How you buy it

How you buy it
CriterionValorBrainLetta
Free planyessource↗One free projectyessource↗USD 0/mo; 3 agents, BYOK
Paid plan purchasable online today to weighnosource↗Contracted directly with the team; online subscription soonyessource↗Buy credits and upgrade online
Published price (model and entry)yessource↗Fixed monthly by number of people; nothing per answeryessource↗subscription+usage; Pro USD 20/mo
Vendor-published benchmarkyessource↗LoCoMo 96.58% per-session recall@10 · BEAM-100K average 75.5% (binary 80.8%)yessource↗LoCoMo 74.0% (GPT-4o mini, Filesystem)

See every alternative

Updated Nov 22, 2026. Found a wrong cell? Write to valorbrain@valor.digital and we fix it with the source.

Why ValorBrain

What a team gets by using ValorBrain as the memory of its people and agents, in six differentiators you can check at the source.

  1. Memory with a current version and authority

    When a fact changes, the new value takes over without erasing the history, and every value says who asserted it. The agent gets the version in force today, even when an older one looks more like the question.

    Proof ↗ Keyed facts win over old prose; authority human > designated > import > agent; the value can be read as of any date (as_of).

  2. Documents and vectors in Brazil, LGPD and on-prem on Enterprise

    DPA and subprocessors published, legal basis and data-subject rights in the Privacy Policy. On Enterprise, on-prem inside your perimeter or a dedicated instance, under contract.

    Proof ↗ Documents, vectors and the database stay in Brazil. Answer generation uses a model provider outside the country, and the encrypted backup lives in the cloud (Google Drive).

  3. Per-source permissions and an access log

    Read grants per business domain for a person, role or agent, with validity and reason. Every read records who read it, through which agent and which version.

    Proof ↗ Controls verified in production on 2026-09-24: per-source ACL (enabled per tenant), read log and an append-only mutation trail with a hash chain.

  4. MCP with real OAuth, and hooks in the main agents

    One hosted MCP endpoint: Claude.ai connects through OAuth, with no token to paste. Hooks bring the relevant memory to the agent at session start and on every prompt (where the agent allows it), without it having to remember to search.

    Proof ↗ OAuth 2.1 with registration_endpoint (DCR) · hooks: Claude Code, Codex, Gemini CLI, Kiro, Cursor.

  5. Ranking without an LLM, 0 calls per search

    BM25 and vectors fused by RRF and reranked by a cross-encoder. An LLM only expands the query; it never decides what comes back. Predictable, auditable search with no model cost in ranking.

    Proof ↗ BEAM-100K, end-to-end answers: average 75.5% (binary 80.8%) (run 2026-08-23, reader GLM-5.3 Flash (Ox Alpha), effort=max). Method and limits at /benchmarks.

  6. Flat price per person

    Agents don't count as seats and nothing is billed per answer or per document, so the team uses it without watching a meter. Start free, no card.

    Proof ↗ Solo US$39, Core US$99, Operations US$199 per month; annual = 12 months for the price of 10.

Prove it in your own workflow

Start on Free in minutes, or talk to us about a team or Enterprise plan.

Rather talk directly?WhatsApp

Start free

Still deciding?

Minutes to set up. Start with one free project and compare on your own stack.

Start with a project