Every Session Starts Cold
Each agent starts from its own prompt context. What one agent decided is invisible to others unless explicitly passed as a message — and message passing breaks when agents run in parallel.
Three Agents.
One Shared Memory
Agents work better when they actually know who they're helping, every session, every time.
Problems
The Cost of Stateless AI
Without persistent shared memory, every agent starts from scratch, duplicates work and loses valuable context between conversations.
No Follow-Through on Tasks
Agents run in isolation. If a Planner deprecates a module, the Coder never sees it and rebuilds from scratch. Conflict detection happens after both agents have already finished their work.
Token Budgets Overflow
Passing each agent's full output to the next as a prompt input fills context windows fast. A research agent running for 10 minutes produces more text than most models can receive.
Features
The Memory Does the Merging
Current sessionJan 15 · 10:05
Last time you mentioned preferring Python and an open auth-token model. Pick up where you left off?
Yes please. Thanks for remembering!
Previous sessionJan 4 · 09:30
I only work in Python, and there's this auth-token recall I'm trying to fix.
PythonAuth-token: open
Every finding is an episode
Agents append raw events to one shared subject. Episodes are content-hashed and immutable. The full provenance trail lives in the log.
01
POST /v1/episodes
Append raw, content-hashed events.
02
POST /v1/memories/compile
Turn episodes into typed memories with confidence and provenance.
idempotentno GPUno vector DB
03
POST /v1/context
Return a ranked, token-bounded context bundle.
Ingest. Compile. Use.
Three endpoints handle the loop. Compile is idempotent. Run it again and again on the same subject. Same query, same bytes.
Bloomberg
3.5%
stale memory
TechCrunch
2.9%
fresh memory
↓
supersede · auto
0.78 ≥ 0.60mem_01 overlaps mem_02
mem_01 SUPERSEDED
provenance + p_03, p_06
Works across multiple agents
Any agent in a pipeline reads and writes the same shared memory, with zero reruns on resume.
What is Stripe's current processing fee?
Bloomberg · 3.5% + 35c
Excluded from context
TechCrunch · 2.9% + 30c
Corroborated by Earnings · p_07
Synthesis
1.2k tok
Stripe charges 2.9% + 30c per transaction.
Personalizes at scale
Adapts responses to the individual's stack, history, and goals, not just the question asked.
Reference Builds
A Shared Memory Layer for Multi-Agent Pipelines
01
Three agents. One memory. Conflicts resolved automatically.
The Planner, Coder, and Reviewer all write to and read from the same Statewave subject. A decision written by any agent is immediately available to every other agent before they act — without any explicit message passing.
01
Bloomberg Agent
Reads market data and extracts price signals.
POST /v1/episodes
02
TechCrunch Agent
Reads the news feed and extracts fresh headlines.
POST /v1/episodes
03
Earnings Agent
Reads financials and extracts EPS and revenue.
POST /v1/episodes
shared subject · market-intel
02
Answers from active memory only.
Context bundles are ranked and token-bounded. Each agent receives only the memories most relevant to its task — not the full episode log. High-signal facts surface first; low-signal filler is dropped.
Active context
8 facts · ~905 tokmarket-intel · /v1/context
User
What is Stripe's current pricing?
Synthesis
Stripe charges 2.9% + 30¢ per transaction.
TechCrunchEarningsp_07
Never returned
Bloomberg · Stripe · 3.5% + 35¢
retired by compiler · jaccard 0.78
03
Kill any agent mid-run. The pipeline doesn't restart.
Each agent's work is durably persisted as episodes the moment it's written. If the Writer is killed mid-run, the Researcher and Critic do not re-run — the Writer picks up from the last compiled context and completes its work.
T+0s
Pipeline started
Three agents launch concurrently against subject market-intel.
T+9s
Earnings killed
^CBloomberg and TechCrunch findings are already compiled and cached.
T+11s
Earnings resumed
resumeReads cached context from /v1/timeline. Upstream agents do not rerun.
Pipeline recovered
0 rerunsActive Memory
The Wrong Fact Never Reaches the LLM
Statewave's compiler retires stale memories before they reach the prompt. Token usage drops, and only facts that still hold are sent to the model.
- Only active memories returned by /v1/context
- Token ceiling enforced before recall
- No GPU. No vector database. No merge logic.
Paste-everything prompt
Tokens to LLM4,000 / session
Both Stripe rates are included. The LLM has to guess which one is current.
Statewave context bundle
Tokens used · average800 / session
Only active memories are included. Bloomberg's stale rate never reaches the prompt.
Memory connected
JavaScript
import StateClient from '@statewave/sdk';
const client = new StateClient({
apiKey: 'your-api-key'
});Python
import os from statewave import StateClient os.environ["STATEWAVE_API_KEY"] = "your-api-key" client = StateClient()
Developer API
Three Endpoints. That's the Core Loop.
Statewave drops into your existing stack with almost no integration work. Connect your agents, persist memory and start shipping in minutes.
- Sub-150ms latency for real-time experiences
- SOC-2 and HIPAA compliant with secure storage
- Compatible with every AI framework and tool.
Why Statewave
Purpose-Built Memory Layer for AI Agents
Built specifically for long-running AI systems where persistent, shared memory is part of the architecture—not an afterthought.
Typed, Ranked Memory
Every memory has a type, confidence score and provenance. Conflicts are resolved automatically, and superseded memories never surface again.
Multi-User by Default
Subjects isolate memory per tenant, user or run. One Statewave instance scales to thousands of concurrent agents with zero cross-contamination.
Token Budget Control
Set max_tokens on every context call. The ranked bundle always fits within budget, with the highest-signal memories first.
Durable Across Sessions
Episodes are immutable and append-only. Restart pipelines, resume mid-run or replay failed agents without losing shared state.
Full Audit Trail
Every episode is timestamped and can include a caller_id. Reconstruct any execution through GET /v1/timeline.
Sub-50ms Recall
Compiled memories are pre-ranked for instant retrieval. Context assembly is a single read, not a vector search.
FAQ
Frequently asked questions
How do several agents share memory without passing messages?
They read and write one subject. Each agent appends its findings with POST /v1/episodes, the compiler turns those into typed memories, and every other agent calls POST /v1/context before it acts — so a decision is visible to the whole fleet without a message being routed anywhere. Message passing assumes turn order; parallel agents have none, which is why the shared store is the coordination point rather than the channel.
What happens when two agents record conflicting facts?
The compiler resolves the overlap instead of leaving both in the prompt. Overlapping memories are marked superseded, and /v1/context returns active memories only — so a stale figure never reaches the model for it to guess between. Each entry keeps its provenance: source episode IDs, confidence score, and supersession state, so the retired fact stays auditable even though it is no longer retrievable as current.
What happens if an agent crashes mid-run?
Nothing upstream re-runs. Every agent’s work is durably persisted as episodes the moment it is written, so a killed agent restarts from the last compiled context while its peers’ findings stay cached. GET /v1/timeline reconstructs the chronological chain for the run — what each agent knew when it acted, and what it wrote — which is also how you audit a pipeline afterwards instead of reading logs.
How do I stop shared memory from blowing the context window?
Set max_tokens on every context call. The bundle is ranked first and packed to that ceiling, highest-signal memories first, so each agent receives the slice relevant to its task rather than the full episode log. The budget is enforced before recall, not by truncating a prompt afterwards — the benchmark harness runs the same path at 512, 1,024, 2,048, and 4,096 tokens.
Answers last checked against the Statewave docs and repositories on .
START BUILDING
Give your AI system memory
Persistent memory for LLMs with sharper context, leaner prompts, and conversations that stay personal to every subject.