Hydra-MCP — Cryptographic Memory ProtocolIETF-TRACK

AI Agent Memory · Tamper-Evident Timestamps · Long-Horizon Amnesia Zero · OpenTTT MCP Server

Tamper-evident timestamps sealed into AI agent memory — 100% state recovery (100/100 measured) after context compaction.

100%
Long-Horizon Amnesia Zero recovery rate (100/100 after server restart)
1,000
causal chain verified (chain_broken=False)
2,056
req/s pot_generate (500VU, p99 671ms, error 0%)
2,589
req/s pot_query (1000VU, p99 372ms)
zero
embedding cost — O(1) exact recall, no vector ops
2 min
npx retrofit — no server changes required
What It Is

Compaction becomes sleep, not amnesia.

AI agents forget their work on every context compaction. Letta, LangGraph, Mem0 lean on vector embeddings — fuzzy recall, no guarantee of deterministic state recovery after a restart.

Hydra-MCP (OpenTTT) works at a different layer. It seals every event with a TTTPS temporal proof (PoT), links them in a causal chain, and looks them up in O(1) by SHA-3 hash. After a restart, the entire chain rebuilds from the Redis DAG — cryptographically verified.

Key Differentiators

Deterministic, cryptographic, standards-track.

  • · OpenTTT MCP Server — instant launch from Claude / GPT / Cursor via npx @helm-protocol/ttt-mcp@0.3.0
  • · 8 MCP Tools — pot_generate · pot_query · pot_checkpoint · pot_verify · pot_status · pot_graph · pot_chain · pot_evict
  • · Long-Horizon Amnesia Zero — 1,000 causal chains verified, 100% recovery rate after server restart
  • · SHA-3-based O(1) exact recall — zero embedding cost. No vector DB. No fuzzy search.
  • · IETF-track I-D (draft-03) — based on TTTPS (draft-helmprotocol-tttps-03). Purpose-built for Byzantine-fault-tolerant AI audit.
  • · Offline operation — automatic fallback to local-clock without a Roughtime server
  • · Persistent signing_key — event recovery unaffected by server restart; pub_key remains unchanged
MCP Tools (8)

Works with any MCP-compatible client.

Claude Code, GPT-4o, Cursor — works instantly with any MCP-compatible client.

ToolDescriptionLatency
pot_generateSeal a new event with a TTTPS temporal proofp50 35ms / p99 671ms (500VU k6, GCP CPU)
pot_queryO(1) event lookup via SHA-3 exact hashp50 31ms / p99 372ms (1000VU k6, GCP CPU)
pot_checkpointSeal the current chain state as a checkpoint (240 epochs)<50ms (target, GCP CPU)
pot_verifyVerify event integrity — tamper detection<10ms (target, GCP CPU)
pot_statusQuery current chain state, epoch, and connected event count<5ms (target, GCP CPU)
pot_graphQuery causal chain DAG (reverse index O(1))<20ms (target, GCP CPU)
pot_chainLink event IDs into a causal chain<15ms (target, GCP CPU)
pot_evictSafely remove stale events (evictedEventIds tracked)<10ms (target, GCP CPU)
Long-Horizon Amnesia Zero — Verification Results

100/100 recovery, measured.

Measured on: GCP CPU · Redis DAG rebuild · recovery verified after server restart (in-memory reset).

100/100
Event recovery rate after server restart
1,000
Consecutive causal chain verifications (chain_broken=False)
2⁻²⁵⁶
SHA-3 hash collision probability — exact lookup basis
+5
Checkpoints added (240×N epochs)
Test ScenarioResultCondition
Event recovery after server restart100/100 PASSRedis AOF + DAG rebuild (including in-memory reset)
1,000 consecutive causal chain continuitychain_broken=Falsepot_generate × 1,000 sequential, then pot_graph verification
Recovery after context compaction100% deterministicSHA-3 exact hash → Redis DAG re-query
pot_generate high load (500VU)2,056 req/s · error 0%max 6.55s, p99 671ms
pot_query high load (1000VU)2,589 req/s · error 0%p99 372ms, max 578ms
AI Agent Memory — Competitor Comparison

Deterministic recovery, not fuzzy recall.

Benchmark: deterministic state recovery after context compaction.

Tool Context Compaction Survival Embedding Cost Recall Method 2-min Install Cryptographic Verification IETF I-D Offline Operation
Letta / Mem0 per op fuzzy
LangGraph / LangMem Partial* per op semantic
RAG / vector DB per query fuzzy
OpenTTT / ttt-mcp zero O(1) exact

If fuzzy semantic recall is needed, use OpenTTT alongside a vector DB. For deterministic state recovery with Byzantine-fault-tolerant audit chain, OpenTTT is the purpose-built IETF-track implementation (draft-03).

Architecture

A SHA-3 causal chain in Redis.

  • · Temporal Proof (PoT) — Seal event timestamps via Roughtime multi-source (k≥3) or automatic local-clock fallback
  • · SHA-3 causal chain — Each event's hash is linked as the parent of the next event. O(1) exact lookup.
  • · Redis DAG persistence — AOF mode enables full chain rebuild after server restart
  • · Rollup checkpoint — Automatic checkpoint every 240 epochs. Compresses stale events.
  • · MCP interface — Model Context Protocol standard. Works with any AI client.
  • · mode:local — Fully offline operation without network. No Roughtime server required.
Installation — 2 Minutes

One npx command.

No server installation required. Run instantly with a single npx command.

# Run directly via npm / npx
npx @helm-protocol/ttt-mcp@0.3.0

# In Claude Code
claude mcp add ttt -- npx -y @helm-protocol/ttt-mcp@0.3.0

# Or via Claude Desktop settings.json
{
  "mcpServers": {
    "ttt-mcp": {
      "command": "npx",
      "args": ["@helm-protocol/ttt-mcp"]
    }
  }
}

After installation, use immediately in Claude/GPT:

# Generate an event
pot_generate({ "content": "architecture decision: QUIC transport confirmed", "context": "session-001" })

# O(1) exact lookup
pot_query({ "event_id": "sha3-abc123..." })

# Seal a session checkpoint
pot_checkpoint({ "label": "v0.3.0-release" })
One-Click Test — Live REST API

Test instantly, no install.

Test instantly via REST API — no server installation required. REST-available endpoints below.

# 1. Seal a new event (POST /pot/generate)
curl -X POST https://api.kenosian.com/mcp/pot_generate \
  -H "Content-Type: application/json" \
  -d '{"content":"test event","context":"demo"}'

Expected response (pot_generate):

{
  "event_id":   "sha3-7f4a2b...",
  "timestamp":  "2026-05-31T00:00:00.000Z",
  "proof":      "roughtime:...",
  "chain_head": "sha3-prev...",
  "epoch":      240
}
# 2. O(1) exact lookup (GET /pot/query)
curl "https://api.kenosian.com/mcp/pot_query?event_id=sha3-7f4a2b..."
# 3. Chain state and statistics (GET /pot/stats)
curl "https://api.kenosian.com/mcp/pot_stats"
# 4. Trigger chain rollup / checkpoint alias (POST /pot/rollup)
curl -X POST https://api.kenosian.com/mcp/pot_rollup \
  -H "Content-Type: application/json" \
  -d '{}'

Note: pot_checkpoint is MCP-channel only (Claude / GPT / Cursor via ttt-mcp). It is not available as a direct REST endpoint — use pot_rollup for equivalent functionality via REST.

Pricing

Free to enterprise.

TierPriceQuotaAction
Free (OSS) $0 100 events/day · self-hosted → GitHub
Dev $29/mo 5M events/mo · hosted · 1K req/s
Pro $99/mo 20M events/mo · priority · SLA
Team $299/mo 100M events/month (fair-use) · tamper-evident audit export · EU AI Act Art.12 · overage $0.01/1K
Enterprise Contact Custom SLA · SSO · on-prem · MiFIR Art.22c audit chain → Contact

Free tier: 100 events/day via hosted API — no credit card. Overage (Team): $0.01 per 1K calls beyond 20M/month.

Industry policy direction
“Third-party evaluation could be done by a government agency (similar to the FAA) or a set of private organizations that are authorized and inspected by the government.”
Dario Amodei, “Policy on the AI Exponential” (2026-06) · darioamodei.com
How this maps. An operator-independent record of when an MCP tool was invoked, in what order, and what data it touched is what compliance and audit require.
Get Started

Long-Horizon Amnesia Zero: 1,000 causal chains verified + 100% recovery after server restart.

Letta, LangGraph, Mem0 — all forget after context compaction. OpenTTT persists the SHA-3 causal chain in Redis, so compaction becomes sleep. On wake, the chain is restored intact. Cryptographically.

github.com/helm-protocol/openttt  ·  IETF draft-helmprotocol-tttps-03

Trust Mark
TTTPS
Verifiable
WHEN
proposed
“TTTPS-certified” — a trust mark for agent memory (proposed)

Each memory operation an agent performs can carry a verifiable, tamper-rejecting record of when it happened and in what order — timing-integrity evidence that an agent’s recorded actions cannot be silently backdated or reordered. If TTTPS is adopted as the standard, an agent-memory deployment could carry the “TTTPS-certified” mark.

Proposed mark for explanation only. Any “certified” status is conditional (“could / if adopted”) and would operate under the Kenosian root of trust — not a present-day certification program.