PLRE Meta-Learning
Plan. Learn. Reflect. Evolve. Domain-agnostic adaptive learning loops that turn every interaction into agent improvement — without manual prompt tuning.
The problem
Prompt engineering is perpetual hand-tuning. No systematic improvement loop.
Prompt engineering is the modern equivalent of hand-tuning regex: a senior engineer notices an issue, rewrites the prompt, ships it, and repeats. There's no systematic improvement loop. Each prompt change is an unevaluated guess.
For agentic applications this is acute. The agent talks to thousands of users, handles thousands of contexts, but the prompts that drive it evolve only when an engineer manually edits them. The system never learns from its own runs.
How it works
The PLRE cycle — agents that improve from each interaction.
The PLRE Engine runs a Plan → Learn → Reflect → Evolve cycle continuously per agent. The Plan module generates objectives + milestones + strategy. The Learn module executes the strategy and captures outcomes into the memory fabric.
The Reflect module analyzes outcomes against the plan — what drove compliance, what caused dropout, what specific prompts produced bad outputs. The Evolve module updates strategy and prompts with confidence scores; rollback fires automatically on quality regression. The whole cycle is grounded in the memory fabric — calibration accumulates across cycles, agents, and users.
- Plan: agent generates objectives + milestones + strategy for the current context
- Learn: agent executes; user feedback + outcome signals are captured into memory
- Reflect: outcomes are analyzed against the plan; what worked vs what didn't
- Evolve: strategy + prompts update with confidence scores; rollback on regression
- Domain-agnostic: works for coaching, customer support, agentic workflows, code review
- Cycle metadata stored in memory fabric — across cycles, agents accumulate calibration
# PLRE meta-learning loop — runs continuously per learner/agent
from cerebe import Cerebe, PLREEngine
cb = Cerebe(api_key=...)
plre = cb.plre(agent_id="coaching-agent-v2")
# Each PLRE cycle: Plan → Learn → Reflect → Evolve
result = await plre.execute_cycle(
learner_id="user_42",
context={"session": "sess_123", "domain": "physical-therapy"},
)
# {
# "plan": {
# "objective": "Reduce left-hip stiffness over 4-week program",
# "milestones": ["week-1: assessment", "week-2: low-impact mobility", ...],
# "strategy": "progressive-overload-with-recovery",
# },
# "learning_result": {
# "interactions": 7,
# "user_feedback": "++compliance, --pain (3/10 → 1/10)",
# "adaptation_signals": ["prefer-morning-sessions", "responds-to-data-viz"],
# },
# "reflection": {
# "what_worked": "morning slot + data-viz drove compliance",
# "what_didnt": "long sessions caused dropout",
# "insights": ["shorter sessions, more frequent", "always include data viz"],
# },
# "evolution": {
# "strategy_update": "switch to 4x/week 20min sessions",
# "prompt_diff": "+include 7-day mobility chart in every session summary",
# "confidence": 0.84,
# },
# }Get Started
Agents that improve every week.
Plan → Learn → Reflect → Evolve. Domain-agnostic adaptive learning. No more manual prompt tuning. Self-signup and full PLRE docs at cerebe.ai.