See why engineering teams replace probabilistic vector databases, LLM write-wrappers, and static AST wiki-graphs with Memoreal's deterministic memory engine (100.0% on LongMemEval in 12.9 ms, 70%+ Claude Code token savings, and $0 LLM write cost).
Replace Mem0's 1–2 paid LLM calls per write and probabilistic vector extraction with 100.0% (500/500) LongMemEval accuracy in 12.9 ms and $0 write cost.
Compare Memoreal's 12.9 ms deterministic memory engine (500/500 LongMemEval) against Zep/Graphiti's async LLM extraction pipeline (451/500 · 90.2% at 104–162 ms).
Keep Claude Code, Cursor, Windsurf, and OpenCode as your agent runtime while adding 12.9 ms deterministic memory (100% LongMemEval vs. Letta's 74.0% LOCOMO).
Skip multi-second Python LLM graph-building jobs (cognee.cognify) and get synchronous <15 ms writes, 12.9 ms recall, and 30 MCP tools out of the box.
Go beyond consumer bookmarking and RAG wrappers with deterministic Supersedes belief revision, 30 code-aware MCP tools, and 100.0% LongMemEval accuracy.
Stop Claude Code from re-reading your entire repository every session. Get 70%+ token savings (turning a $20 Claude Pro plan into a $100 Max plan) at <20ms Rust latency — without the terminology collisions and git-checkout staleness of static AST/Obsidian wiki-graphs.
Benchmark charts are nice; not shipping broken PRs is better. Click through 5 everyday engineering workflows where vector databases, rerankers, and LLM memory wrappers hallucinate — and see how Memoreal nails the exact answer in 12.9 ms for $0.
Top-1 Match (0.91 cosine): "Evaluated Kafka for billing webhooks..." Agent action: Scaffolds 400 lines of dead Kafka consumer boilerplate.
Cosine similarity and cross-encoder rerankers are blind to negation. They match "Kafka" + "billing webhooks", rank the canceled POC #1, and send Claude Code or Cursor down the wrong architecture for an hour.
Active State: PostgreSQL LISTEN/NOTIFY (Kafka POC marked rejected). Agent action: Wires retry handler directly into existing Postgres listener.
Negation-aware state tracking and Supersedes belief revision know the difference between "we run Kafka" and "we killed Kafka." Rejected RFCs stay buried — in 12.9 ms with 0 LLM tokens.
Retrieved context: DynamoDB us-east-1 + 4 unrelated vendor specs. Raising top_k (5 → 20): Dumps 4x more third-party noise into the prompt.
Detailed vendor docs pack denser technical keywords and outrank your own architecture notes. Worse, widening top_k from 5 to 20 drags in even more external threads — multiplying false attribution instead of fixing it.
First-Party Core API DB: PostgreSQL 16 (europe-west1). External vendor context (Stripe/DynamoDB) isolated from system ownership.
First-party ownership resolution, provenance trust chains, and strict namespace isolation separate your production stack from external vendor docs automatically — zero prompt bloat, zero wrong-database PRs.
LLM write summary ($): "Project auth uses passkeys." Agent action: Breaks working session middleware chasing nonexistent passkey routes.
LLM write-wrappers charge you for 1–2 external LLM calls on every turn just so a summarizer prompt can strip away "maybe in Q4" and overwrite verified production state with an unstarted brainstorm.
Core Fact: Express session auth (active). Quarantined Idea (Episodic): Q4 passkeys brainstorm — unconfirmed, never promoted.
Speculative-statement guardrails and 4-Tier memory promotion (Buffer → Episodic → Semantic → Core) keep unconfirmed ideas from overwriting verified production facts — in <15 ms at $0 write cost.
Core block full (2,000 chars) → LLM rewrites summary. Evicted: org_id tenant isolation rule. Kept: yesterday's button border radius.
Agents that iteratively rewrite a fixed-size scratchpad (Letta core blocks or auto-summarized MEMORY.md files) compress away early architectural invariants to make room for recent UI chatter.
<memory_context> (SessionStart & PreCompact · 12.9 ms): [Core 0.95] Never run git rebase; always scope tenant queries by org_id.
Core and high-salience Semantic rules persist permanently without summary-rewrite drift, injecting a deterministic 4,096-token <memory_context> on every SessionStart and PreCompact.
Retrieved chunks: "...worker health check port 8080 (512 MB RAM)..." (missing +4 deploy). Agent answer: Hallucinates 512 or guesses 2 from partial chunks.
Rerankers boost unrelated log lines ("512", "8080") because they sit next to server keywords, while top-k truncation drops your initial +4 provisioning event — guaranteeing a wrong count.
State transitions: +4 provisioned → +2 scaled → -3 drained = 3 active worker nodes. Unrelated port 8080 & 512 MB metrics filtered automatically.
Deterministic stateful math and category hierarchies track every +4, +2, and -3 delta across sessions while ignoring unrelated port and memory metrics — verified 500/500 (100.0%) on LongMemEval.
Why engineering teams replace LLM-wrapper memory services and vector databases with Memoreal's deterministic memory engine.
| Capability | Memoreal | Hindsight | Mem0 | Zep / Graphiti | Cognee | Letta (MemGPT) |
|---|---|---|---|---|---|---|
| LongMemEval Accuracy | 100.0% (500 / 500) | 94.6% (LongMemEval-S) | 93.4% (claimed) | 90.2% (451 / 500) | Not published | 74.0% (LOCOMO) |
| Multilingual / Hebrew LongMemEval | 100.0% (102 / 102 · 15 langs) | English-centric | LLM prompt dependent | English-centric | LLM prompt dependent | LLM prompt dependent |
| LLM Tokens Spent per Write | 0 tokens ($0) | LLM extraction / reflect() | 1–2 LLM calls / write | Async LLM extraction | Python ECL LLM pipeline | Agent tool-call tokens |
| Recall Latency | 12.9 ms | 50–350 ms (Postgres) | 400–1,800 ms | 104 ms p50 / 162 ms p95 | 250–1,400 ms | 500–2,500 ms |
| Deterministic Belief Revision | Automatic Supersedes revision | Entity consolidation | Probabilistic LLM prompt | Temporal window (t_invalid) | Graph re-indexing | Agent file overwrite |
| Code Graph & MCP Tools | 30 MCP tools + 3 resources | Memory bank tools | 11 text memory tools | Graph search tools | 14 MCP tools (codify) | Agent harness only |
One installer automatically configures Claude Code, Cursor, Windsurf, OpenCode, Claude Desktop, OpenAI Codex CLI, Gemini CLI, VS Code Copilot, Zed, Cline, Roo Code, Kilo Code, Continue.dev, Goose, and Amp.
Includes 8 automatic lifecycle hooks (SessionStart, PreCompact, PostCompact, UserPromptSubmit, PostToolUse, Stop), 4 slash commands (/memoreal-sync, /memoreal-sync-code, /memoreal-wrap-up, /memoreal-doctor), 30 MCP tools, and 3 MCP resources (@memory://rules).
curl -fsSL https://memoreal.kolivri.com/downloads/install.sh | MEMOREAL_KEY=mem_dev_your_key_here bash
@kolivri/memoreal-mcp · stdio MCP bridge for Claude Code, Cursor, OpenCode & Claude Desktop)
claude mcp add memoreal -e MEMOREAL_API_KEY="mem_dev_your_key_here" -e MEMOREAL_API_URL="https://memoreal.kolivri.com" -- npx -y @kolivri/memoreal-mcp
~/.cursor/mcp.json · auto-configured by installer, 1-click button above, or paste JSON)
{
"mcpServers": {
"memoreal": {
"type": "http",
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.codeium/windsurf/mcp_config.json · auto-configured by installer or paste JSON)
{
"mcpServers": {
"memoreal": {
"type": "http",
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.codex/mcp.json + ~/.codex/AGENTS.md)
curl -fsSL https://memoreal.kolivri.com/downloads/install.sh | MEMOREAL_KEY=mem_dev_your_key_here bash
~/.config/opencode/plugins/memoreal-hooks.js + opencode.json)
curl -fsSL https://memoreal.kolivri.com/downloads/install.sh | MEMOREAL_KEY=mem_dev_your_key_here bash
claude_desktop_config.json · auto-configured by installer)
{
"mcpServers": {
"memoreal": {
"type": "http",
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.vscode/mcp.json · auto-configured by installer or 1-click button)
{
"mcpServers": {
"memoreal": {
"type": "http",
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.config/zed/settings.json · auto-configured by installer)
{
"context_servers": {
"memoreal": {
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.cline/mcp_settings.json), Roo Code (~/.roo/mcp.json) & Kilo Code (~/.kilocode/mcp.json)
{
"mcpServers": {
"memoreal": {
"type": "http",
"url": "https://memoreal.kolivri.com/v1/mcp",
"headers": { "Authorization": "Bearer mem_dev_your_key_here" }
}
}
}
~/.gemini/settings.json), Goose (~/.config/goose/mcp.json), Amp & Continue.dev
curl -fsSL https://memoreal.kolivri.com/downloads/install.sh | MEMOREAL_KEY=mem_dev_your_key_here bash
$env:MEMOREAL_KEY="mem_dev_your_key_here"; irm https://memoreal.kolivri.com/downloads/install.ps1 | iex
https://memoreal.kolivri.com/v1/memories/context)
curl -fsSL --proto '=https' --tlsv1.2 -X POST https://memoreal.kolivri.com/v1/memories/context \
-H "Authorization: Bearer mem_dev_your_key_here" \
-H "Content-Type: application/json" \
-d '{"namespace":"my-project","query":"architecture decisions and active tasks","token_budget":4096}'
f4842aad57fe07ecb9e881b48b2b648504f4a462abcbb16eb5e111bd24704229 · (.zip): 20a2f5af3e31a8cf32a41f5056f6f2e9771a7ae98379bdade5d261f1df9132f0
SessionStart and restores working state after context compaction.