AgentRealm · Research

Your agent remembers. Nobody governs what it kept.

An always-on agent doesn’t just recall — it carries durable state that causes actions later: task ledgers, permissions, credentials, commitments, triggers, audit records. This survey codes 435 works and finds a field fluent in putting state in, and nearly silent on getting it out, revoking it, or undoing what it did.

Read the paper

arXiv:2606.30306 · cs.MA · Ding · Nannapaneni · Liu · Zhang · 435 works, coded four ways

01 — what counts as always-on

Always-on is not about running constantly.

The survey’s definition: an agent is always-on if its policy at time t depends on state accumulated before t — beyond the current prompt or task instance. The defining property is persistence, not continuous execution.

It’s written to be checked against a system rather than asserted of it. Three conditions classify it.

GAP
GAP
agent #1 — same instance, survives every restart
Session 1
monday
Session 2
wednesday
Session 3
next month
"repo uses tabs"
this record is what acts
rewrites your file to match
01
persistent identity
survives restarts
02
agent-owned durable state
outlives the session
03
it acts on that state
the loop closes
Always-on
A convention written weeks ago is still authoritative, and it is what shapes the action taken now.
the fourth condition — an obligation, not a test
Temporal accountability

If the retained state is consequential, the system owes an answer when someone asks why it acted. Four questions, and most systems can answer none of them.

audit at t — why did you act?
What did you know?
not recorded
When did you learn it?
not recorded
Why did that license this action?
not recorded
If it was poisoned, what do you undo?
not recorded

The first three conditions classify a system. This fourth one is an obligation — and it is what the rest of this survey measures.

PERSISTENTSTATE
Memory
Task ledgers
Permissions
Credentials
Provenance
Social state
Triggers
Commitments
Memory
each satellite is persistent state
1 / 8

Memory is one satellite. Persistent state is the planet.

Persistent state is every durable, control-relevant record that crosses the session boundary with the agent — not just the retrievable text most prior work studies. They fail together, so they have to be governed together.

Memory
Retrievable memory — the object most prior work studies, and only one entry on this list.
02 — six axes of persistent state

Every durable record has six dimensions. The field measures the safe ones.

Each axis names a question a stored item must answer before it is allowed to move the world. Pillar height is how many of the 435 works touch that axis.

AUTHORITY
72 works
SCOPE
uncounted
MUTABILITY
160 works
PROVENANCE
153 works
RECOVERABILITY
112 works
ACTIONABILITY
uncounted
AUTHORITY
step through the six axes
1 / 6
axis
AUTHORITY
Who permits this to act?

The line between a fact the agent holds and one it is licensed to act on. The rarest axis in the corpus — and the first thing poisoning, privilege drift and cross-user leakage attack.

coverage · 72 of 435 — rarest

The survey reports per-axis counts for four of the six. Scope and actionability are named and argued but not separately counted — drawn as dashed spokes rather than invented numbers.

03 — the state lifecycle

The loop has two arcs. The field only built one.

State is observed, written, validated, organized, retrieved, acted on — then updated, forgotten, audited, rolled back. Sphere size is how many of the 435 works exercise that stage.

FORWARD ARCaccumulate & useRETURN ARCgovern & recover
observe
write
validate
organize
retrieve
act
update
forget
audit
rollback
68
200
87
128
269
141
127
66
88
27
observe
step through the ten stages
1 / 10
observe
68 / 435
forward arc

No governed criterion for what deserves to become durable. The write captures the value and drops who authorized it.

representative work · Anticipatory write rehearsal (Guo et al., 2026); user-driven write (Huang et al., 2023)
five invariants the loop should hold
01Authority monotonicity — permission may narrow, never silently widen.
02Scope non-expansion — a merge must not leak across principals.
03Deletion propagation — erasure reaches derived copies.
04Provenance preservation — consolidation keeps the source.
05Rollback traceability — every effect carries a handle home.

Three of the five live on the return arc — the arc the corpus shows is least implemented.

04 — what breaks

Persistence invents its own failure modes.

None of these are hallucinations. Each is a durable record doing exactly what it was stored to do — at the wrong time, for the wrong principal, with no way back. Each names the stage it breaks at and the invariant it violates.

tool output
written to memory
session 2
inherits it
session 9
still inherits it
attacker long gone
at write · breaks provenance preservation
Write poisoning

One untrusted line of tool output is promoted into durable memory. It doesn't need to be retrieved today — it waits, and every future session inherits it, long after the attacker's session ended.

query
salient · wrong
salient · wrong
irrelevant
context
budget
the record you needed
crowded out
at retrieve · breaks scope non-expansion
Retrieval distraction

More recall is not more competence. Salient-but-wrong episodes crowd out current evidence and drag tool choice off course — in several studies a memoryless baseline simply wins.

permission granted
revoked at t₂
trigger stored at t₁
GATE OPEN
fires at t₃
nothing marks the record as expired
at act · breaks authority monotonicity
Stale commitment

A trigger stored under a permission that has since been revoked still fires. Nothing marks the record as expired, so stale state quietly overrides fresh observation.

delete ✓
index patched
derived
summary · still holds it
derived
embedding · still holds it
at forget · breaks deletion propagation
Deletion residue

Delete patches one index; the summary and the embedding that already absorbed the value keep it. Decay-based forgetting is worse — downweighted, present, and still retrievable under the right query.

agent A · bad write
shared pool · untyped, unaudited
agent B
agent C
agent D
same premise
at organize · breaks scope non-expansion
Cross-agent propagation

In shared memory one agent's bad write becomes every agent's premise. Publish-subscribe pools are untyped and unaudited, with no revocation semantics and no one owning cleanup.

SNAPSHOT
restore brings back
deleted data
revoked permission
27 of 435 works · the least studied stage
at rollback · breaks rollback traceability
Rollback itself

The cruel one. Restoring a snapshot can resurrect deleted data and re-grant revoked permission. The operation most likely to re-introduce corruption is the least studied of all.

05 — the point

A bigger context window fixes none of this.

one persistent-state record8 fields
value
"deploys need Ana’s approval"
authority
✕ not represented
scope
✕ not represented
provenance
✕ not represented
effect handle
✕ not represented
actionability
✕ not represented
logical time
✕ not represented
retention
✕ not represented

Value is the only field most memory systems represent. Discard the envelope and nothing can recover it — the record cannot say whether it is still authorized, or what it has already caused.

01
The missing thing is a field, not a fact.

Recall and governance are different capabilities. A text store has nowhere to put a permission epoch, a deletion ledger or a rollback handle — so no amount of reading solves it.

02
Score mutations, not answers.

The survey proposes AOEP-v0: a pilot contract that scores state mutation and recovery obligations instead of answer quality, split so a system that stores nothing cannot look safe. Proposed, not shipped.

03
Borrow from fields that solved it.

Databases, distributed systems, formal methods, capability security, machine unlearning. Transactions, compensation, revocable grants, verified erasure — old work the agent stack has not imported yet.

Always-On Agents: A Survey of Persistent Memory, State, and Governance in LLM Agents
Tianyu Ding · Aditya Nannapaneni · Bingfan Liu · Ling Zhang · arXiv:2606.30306 [cs.MA]

All counts are the survey’s own coding of a 435-work corpus — a scoped map, not an exhaustive census of the field. Last updated: August 19, 2026.