A VB Pulse survey this June found that 57% of enterprises had traced a confidently mistaken agent reply again to lacking or inconsistent context — the newest signal of how central context has turn out to be as to whether AI brokers might be trusted to behave on their very own.
Many of the fixes to date have solved a narrower model of that downside: one agent remembering extra, in a single session. What's been lacking is a means for a workforce of brokers to attract on the identical context directly, and that hole is the place a more moderen downside is surfacing. As soon as an agent's context is shared throughout an entire workforce, a mistaken reality doesn't price one particular person a repeated clarification. It prices the entire workforce.
Tencent's reply to that hole is Agent Memory, an open-source venture the workforce stated grew out of six months spent fixing a narrower downside: brokers shedding context in lengthy periods. A part of that system is a persona layer, a steady, distilled image of who a person is and the way they work, constructed up over many conversations moderately than reconstructed every time. On Tencent's personal benchmark for whether or not an agent nonetheless applies that image appropriately after prolonged use, accuracy rose from 48% to 76%, a 59% relative enchancment, as soon as the persona layer was added. This week, Tencent prolonged that venture with the beta launch of Staff Reminiscence, which opens the identical strategy as much as an entire workforce as an alternative of 1 agent. Tencent stated the repo hit No. 1 on GitHub's TypeScript trending list this week.
Brokers on a workforce can now learn from a shared reminiscence hub as an alternative of preserving separate, siloed context, ruled by means of an entry management layer that determines who can learn what.
What Staff Reminiscence truly does
The core thought is a shared hub moderately than a shared immediate. As an alternative of pasting one giant context block into each agent's window, Staff Reminiscence registers 4 sorts of reusable belongings and equips every agent with solely those it wants.
-
Chat Reminiscence. Retains preferences, details, selections, and interplay historical past, distilled by means of 4 layers, from uncooked dialog as much as a steady long-term persona, so an agent doesn’t must be reintroduced to a person it has already labored with.
-
Ability. Captures procedures pulled from accomplished work, versioned and reviewed earlier than they’re shared moderately than dropped right into a folder as-is.
-
LLM-Wiki. Turns paperwork and specs into structured, linked pages.
-
Code-Graph. Indexes a codebase's symbols, information, and name relationships so an agent can test what a change may have an effect on earlier than making it.
Tencent's documentation attracts the excellence immediately: "RAG solutions 'what might be discovered?' Staff Reminiscence additionally solutions 'who can use it, which model is legitimate, and which Agent ought to obtain it.'"
In observe, that's what Tencent calls an "Agent Loadout": a Scout agent doing analysis might be geared up with market analysis and aggressive evaluation belongings, whereas a Builder agent will get the code graph and product docs it wants as an alternative, moderately than each agent gaining access to all the pieces.
Which belongings an agent will get geared up with is ruled by means of 4 visibility tiers:
-
Personal. Readable solely by the asset's proprietor.
-
Staff. Readable by anybody on the workforce.
-
Restricted. Gated by person, position, or agent-level entry management.
-
Agent. Outfitted to 1 particular agent inside a workforce.
New belongings default to non-public, so sharing needs to be a deliberate motion moderately than one thing that occurs routinely.
What occurs when a reminiscence is mistaken
That entry mannequin solutions an actual query, who’s allowed to learn a given reminiscence asset. It doesn’t reply a second one, which is what occurs as soon as a reminiscence asset seems to be mistaken. Tencent's personal documentation lays out possession, versioning, and standing monitoring for every asset, however nothing within the documentation describes a correction or expiry course of for a incontrovertible fact that's already been learn and reused by different brokers on a workforce, or a solution to resolve it when two brokers' recollections of the identical factor disagree.
That hole is what practitioners flagged inside hours of the launch publish.
"Shared reminiscence makes the write path the attention-grabbing downside. Retrieval will get many of the consideration, however a mistaken reality written as soon as now propagates to each teammate's agent as an alternative of simply yours. Curious how the governance layer handles correction and expiry," Blake Murphy wrote on X.
The priority wasn't solely about fixing a foul reality after the actual fact. It was concerning the determination to depart one thing out of the file within the first place. "the ruled half is the laborious half. as soon as teammates' brokers can learn one another's context, somebody has to resolve what by no means will get written down," Virgil Maro wrote on X.
Others pushed additional into what occurs as soon as two brokers' recollections actively contradict one another, not simply go stale.
"The Code-Graph plus LLM-Wiki break up is the correct name. The half I'd wish to see benchmarked: in shared mode, whose reminiscence wins when two teammates' brokers have written contradicting details about the identical module? Single-agent reminiscence drifts slowly. Shared reminiscence drifts quick, as a result of one stale write propagates to individuals who by no means noticed the session that produced it," Austin Green wrote on X.
The response wasn't uniformly important. "Fascinating shift: making reminiscence a shared service turns brokers into an actual workforce moderately than remoted bots. Governance would be the trickiest half, particularly when details battle," Moez Zhioua wrote on X.
None of those are edge circumstances particular to Tencent's implementation. A March 2026 paper on manufacturing multi-agent reminiscence structure, "Governed Memory: A Production Architecture for Multi-Agent Workflows," revealed independently of any single vendor, identifies governance fragmentation and silent high quality degradation with out suggestions loops as structural dangers in shared multi-agent reminiscence typically. The sample the paper describes matches what the commenters above pointed at immediately: a mistaken reality in a single-agent reminiscence system prices one person a repeated correction, whereas the identical mistaken reality in a shared, team-wide reminiscence system propagates to each agent that inherited it earlier than anybody catches it.
How Staff Reminiscence compares
AI agent reminiscence work in 2026 has principally centered on a single agent remembering extra, in a single session, about one person: LangChain's LangMem SDK, Google's Always On Memory Agent, and Anthropic's work contained in the Claude Agent SDK all work this fashion. A unique line of labor has centered on giving brokers entry to a shared mannequin of enterprise information. VB's personal June survey discovered solely 25% of enterprises had that form of ruled context layer in manufacturing, whereas distributors together with AWS, Couchbase, Oracle, Redis, and Pinecone have all shipped variations of it this 12 months.
Staff Reminiscence's closest present comparability is probably going Asana, which constructed shared memory across a company's AI teammates so an agent doesn't must be re-briefed on context one other agent already has. Asana's CPO described the identical tradeoff Tencent's practitioners at the moment are elevating, an access control system built specifically to stop one agent's memory from leaking into a project another agent isn't cleared to see. Tencent's model is open-source and moveable throughout frameworks moderately than scoped to 1 platform, but it surely's answering a query Asana's workforce already bumped into whereas constructing a closed one.
For groups evaluating this class, the upside is actual: brokers cease relearning what the workforce already is aware of. The tradeoff is simply as actual: one unhealthy write is not contained to 1 agent — it's inherited by each agent that reads from the shared pool, with no correction or expiry course of but in place to catch it.
