# Option A: zero-install (recommended)
uvx agent-sleep-mcp

# Option B: install first, then run the init helper
pip install "agent-sleep[mcp]"
agent-sleep init     # prints the correct config snippet for your platform

{
  "mcpServers": {
    "agent-sleep": {
      "command": "uvx",
      "args": ["agent-sleep-mcp"]
    }
  }
}

"Before we start, check your memory for anything relevant to this task."
"Record that we use pytest fixtures — not unittest — in this project."
"Run a sleep consolidation so you remember today's lessons next session."

# See all memories and rules for the current project
agent-sleep show

# Clear a project's memory (with confirmation prompt)
agent-sleep reset

# Target a specific scope or DB
agent-sleep show --scope my_api --db /path/to/memory.db

          [ ONLINE EXECUTION PHASE ]
            Agent executes tool calls
                       │
                       ▼
┌──────────────────────────────────────────────┐
│  1. EPISODIC RECORDING                       │
│     memory.record_episode(...)               │  Fast, minimal overhead.
│     Records goal, action, outcome, errors.   │  Stores execution events.
└──────────────────────┬───────────────────────┘
                       │
             (Session ends / Agent idle)
                       │
                       ▼
          [ OFFLINE SLEEP CONSOLIDATION ]
┌──────────────────────────────────────────────┐
│  2. SLEEP CONSOLIDATOR (8-Stage Pipeline)    │
│     SleepConsolidator.run(session_id)        │
│                                              │
│     • Priority Replay (prediction error)     │
│     • Deterministic Episodic Distillation    │  Grounding first:
│     • Procedural Recipe Extraction           │  distills facts & lessons
│     • How-Memory Trajectory Abstraction      │  before optional LLM
│     • Behavioral Rule Promotion (seen ≥2x)   │  generalization passes.
│     • Epistemic Status (observed vs verified)│
│     • Episodic Compression over time         │
│     • Self-Competence EMA Tracking           │
└──────────────────────┬───────────────────────┘
                       │
              (Next session / New task)
                       │
                       ▼
          [ ONLINE SELECTIVE RECALL ]
┌──────────────────────────────────────────────┐
│  3. SELECTIVE SEMANTIC RECALL                │
│     memory.recall(new_task)                  │  Pre-computed vector BLOBs.
│     Returns only relevant lessons & rules    │  Prevents prompt dilution.
│     filtered by project scope & relevance.   │
└──────────────────────────────────────────────┘

python benchmarks/agent_eval/runner.py

from agent_sleep import AgentMemory, SleepConsolidator

# 1. Initialize memory scoped to your project/repo
memory = AgentMemory(session_id="session_01", scope="payment_service")

# 2. Record actions and outcomes during your agent's loop
memory.record_episode(
    goal="Refactor payment processor to async",
    action="edit_file('processor.py', ...)",
    outcome="failure",
    failure_reason="SyntaxError: 'await' outside async function",
)

# 3. Trigger sleep consolidation when idle or at session end
consolidator = SleepConsolidator(scope="payment_service")
report = consolidator.run(session_id="session_01")
# -> {'episodes_processed': 1, 'memories_written': 1, 'rules_promoted': 0, ...}

# 4. Next session: recall relevant context before executing
context = memory.recall("Add Stripe webhook handler")
print(context)
# [MEMORY CONTEXT]
# Relevant past experience:
#   ⚠ [LESSON] Caution on task: Refactor payment processor to async:
#     A previous attempt failed: SyntaxError: 'await' outside async function.
# [END MEMORY CONTEXT]

pip install "agent-sleep[mcp]"

pip install "agent-sleep[all]"

pip install git+https://github.com/thevisionhub/agent-sleep.git

git clone https://github.com/thevisionhub/agent-sleep.git
cd agent-sleep
pip install -e ".[all]"

pytest tests/ -v