When Memory Store Listings Returned Half the Rows: Migrating to agent-memory-2026-07-22

The counterintuitive part: realizing I had leaned on substring matching

To sweep across a project and its archive in one call, my audit script did this:

# Before. This picked up both /projects/alpha/... and /projects/alpha_archive/...
memories = client.beta.memory_stores.memories.list(
    memory_store_id=store_id,
    path_prefix="/projects/alpha",   # no trailing slash
    view="basic",
    betas=["managed-agents-2026-04-01"],
)

The old path_prefix was a substring match. "/projects/alpha" is a prefix of "/projects/alpha/decisions/2026-07.md", but it is equally a prefix of "/projects/alpha_archive/old.md". Half-consciously, I had come to rely on that as a convenient way to grab a project and its archive together.

Under agent-memory-2026-07-22, path_prefix now requires a trailing slash and matches by whole segments. Two things follow.

First, "/projects/alpha" without a trailing slash no longer picks up the range I intended, because it does not land on a segment boundary. Second, the correct form "/projects/alpha/" does not include /projects/alpha_archive/, because alpha and alpha_archive are different segments.

So what used to be "one call returns both the live tree and the archive" became "only the live tree comes back." That was the real reason the count looked halved. No error fired, because the path_prefix itself was still syntactically valid; only the matched range had quietly narrowed.

This was the biggest lesson for me. Of all breaking changes, the dangerous ones are not those that throw, but those that silently return something different. The former you always notice. The latter slips past you unless you are watching.

Rewriting the access layer to survive the change

The fix rests on three principles. Normalize the prefix. Do not depend on ordering. Walk the tree explicitly. With those in mind, I rewrote the audit layer as follows. This is lifted from the code that actually runs every morning, with secrets stripped out.

from anthropic import Anthropic
 
client = Anthropic()  # ANTHROPIC_API_KEY comes from the environment
BETAS = ["managed-agents-2026-04-01", "agent-memory-2026-07-22"]
 
 
def normalize_prefix(prefix: str) -> str:
    """A path_prefix must end with a slash. Normalize root to "/"."""
    if not prefix.startswith("/"):
        prefix = "/" + prefix
    if not prefix.endswith("/"):
        prefix = prefix + "/"
    return prefix
 
 
def list_all_memories(store_id: str, prefix: str = "/") -> list[dict]:
    """
    Return every memory under prefix.
    - Do not depend on return order (the caller always sorts)
    - depth accepts only 0/1/omitted, so walk the tree recursively ourselves
    """
    prefix = normalize_prefix(prefix)
    collected: list[dict] = []
    cursor: str | None = None
 
    while True:
        page = client.beta.memory_stores.memories.list(
            memory_store_id=store_id,
            path_prefix=prefix,
            depth=1,          # take only direct children; recurse into subtrees
            view="basic",     # metadata only; retrieve content on demand
            after_id=cursor,
            betas=BETAS,
        )
        for item in page.data:
            if item.type == "memory_prefix":
                # a directory-like node; descend one level
                collected.extend(list_all_memories(store_id, item.path))
            else:
                collected.append({"id": item.id, "path": item.path})
 
        if not getattr(page, "has_more", False):
            break
        cursor = page.last_id
 
    # Server order is not guaranteed, so sort ourselves for stable diffs
    collected.sort(key=lambda m: m["path"])
    return collected

A few notes that live outside the code.

I pinned depth=1 and made the function recurse whenever it hits a memory_prefix (a directory-like node). I used to pass a large depth to grab everything at once, but that value is now a 400. Walking the tree myself frees me from the depth limit and, as a bonus, lets me see the count at each level.

I lean on ordering nowhere. After fetching, I sort by path. A day-over-day diff should be compared as a set, not as a sequence, and written that way it does not break when the server order shifts.

And I do not pull content: view="basic". An audit needs only the path and ID. When I actually need the body, I memories.retrieve a narrowed set. Downloading 100 kB times many files every morning was a waste, both in time and in cost.

Rewriting "the live tree plus the archive" as an explicit intent

The old, substring-driven behavior of grabbing both the live tree and the archive was convenient, but the intent was implicit. I took the migration as a chance to rewrite it as an explicit enumeration.

def list_project_including_archive(store_id: str, project: str) -> list[dict]:
    """Deliberately combine the live tree and archive as two separate prefixes."""
    live = list_all_memories(store_id, f"/projects/{project}/")
    archive = list_all_memories(store_id, f"/projects/{project}_archive/")
    return live + archive

Writing it out showed me just how fragile the old code had been. Once I had more projects named around "alpha," the substring match would have swept in alpha2 and alpha_experimental too. Segment matching did not make things less convenient; it forced my vague intent to become explicit. That is how I have come to see it.

Safe corrections with a content_sha256 precondition

Half of what an audit is for is correcting bad memories, and there is a hazard here that is independent of the version change. Because the agent and I touch the same memories, overwrite conflicts are possible.

memories.update supports optimistic concurrency. Pass the content_sha256 you read, and the update applies only if the stored hash still matches.

def safe_correct(store_id: str, memory_id: str, new_content: str) -> bool:
    current = client.beta.memory_stores.memories.retrieve(
        memory_store_id=store_id, memory_id=memory_id, betas=BETAS,
    )
    try:
        client.beta.memory_stores.memories.update(
            memory_store_id=store_id,
            memory_id=memory_id,
            content=new_content,
            precondition={
                "type": "content_sha256",
                "content_sha256": current.content_sha256,
            },
            betas=BETAS,
        )
        return True
    except Exception:
        # hash mismatch = someone (usually the agent itself) wrote first.
        # do not clobber silently; re-read and decide again
        return False

There really were nights when the agent was appending to the same file. Back when I corrected memories without a precondition, I must have occasionally stomped those appends. Every change is retained as an immutable memory version (memver_...), so you can notice after the fact, but not stomping it is better still.

Counting the blast radius before you migrate

I caught the halving only because I happened to have the previous day's count on hand. Put the other way around: without it, I would not have caught it at all.

Guarding against a repeat turned out to be simpler than expected. As long as you have the list of memory path values, you can measure the blast radius without touching the API. You just apply the old substring rule and the new segment rule to the same set of paths, side by side.

"""prefix_audit.py — count the gap between old substring and new segment matching, offline."""
import json, sys
 
 
def substring_match(paths, prefix):          # the old behavior
    return [p for p in paths if p.startswith(prefix)]
 
 
def segment_match(paths, prefix):            # the new behavior
    if not prefix.endswith("/"):
        return []                            # no trailing slash means no match
    return [p for p in paths if p.startswith(prefix)]
 
 
def prefix_nodes(paths):
    """How many directory-like nodes a depth=1 recursion has to step through."""
    nodes = set()
    for p in paths:
        parts = p.strip("/").split("/")[:-1]
        for i in range(len(parts)):
            nodes.add("/" + "/".join(parts[: i + 1]) + "/")
    return nodes
 
 
if __name__ == "__main__":
    paths = json.load(open(sys.argv[1]))     # paths collected from memories.list
    print(f"total memories: {len(paths)}")
    n = len(prefix_nodes(paths))
    print(f"prefix nodes  : {n}  -> depth=1 recursion needs {n + 1} list calls\n")
    print(f"{'path_prefix':<28}{'substring':>10}{'as-written':>12}{'normalized':>12}{'lost':>7}")
    for pf in sys.argv[2:]:
        old = len(substring_match(paths, pf))
        raw = len(segment_match(paths, pf))
        new = len(segment_match(paths, pf if pf.endswith("/") else pf + "/"))
        print(f"{pf:<28}{old:>10}{raw:>12}{new:>12}{old - new:>7}")

Run it against a tree shaped like mine — six projects, 68 memories — and it prints this.

total memories: 68
prefix nodes  : 15  -> depth=1 recursion needs 16 list calls

path_prefix                  substring  as-written  normalized   lost
/projects/alpha                     34           0          20     14
/projects/alpha/                    20          20          20      0
/projects/beta                      16           0          16      0
/projects/                          68          68          68      0

Three columns carry the meaning. substring is the old match count, as-written is what you get if you send the prefix exactly as it appears in your code, and normalized is what you get after adding the trailing slash.

Lined up like that, two distinct failure modes come into view.

/projects/alpha drops from 34 to 20 — a 41% shortfall. Those 14 are alpha_archive and alpha2 — the places where my intent had been implicit. What I had filed away as convenient behavior may not have been convenient at all; I may simply never have decided where the boundary sat. /projects/beta, by contrast, reads 16 under both substring and normalized. By count alone it looks untouched. Yet as-written is 0. Ship it with the trailing slash still missing and it does not come back partially; it comes back empty.

The rows where lost is 0 are the dangerous ones. They show no difference, so they get filed as "unaffected" and nobody touches them.

Then there is the second line of the output. Fifteen nodes in the tree means a depth=1 recursion turns one list call into sixteen, a 16x increase. For a once-a-morning audit that is noise. For a loop that runs every few minutes it is not. I walk the shallow levels every time and descend into the deeper ones only when the count above them has moved. That estimate of sixteen turns out to be wrong, incidentally; the next section runs it and corrects the number.

Measured: what fifteen recursive calls actually cost, and where naive parallelism jams

Saying "sixteen times as many calls" in the previous section was not enough to act on. If a single call takes 5 ms, sixteen of them still add up to 80 ms. So I measured it.

I stood up a minimal local server that returns the same shape of response as memories.list, with 45 ms of latency injected per call. The tree is exactly the one from the audit output above — 68 memories, 15 prefix nodes. Python 3.10.12, median of three runs.

The first thing the run surfaced was a mistake in my own audit script.

prefix nodes  : 15  -> depth=1 recursion needs 16 list calls
actual walk    : list calls = 15

prefix_nodes() already includes the root /projects/ in its set. Adding n + 1 on top of that counted the root twice. It is a difference of one call, but the reasoning behind the estimate was still wrong. Where the previous section says sixteen, fifteen is the correct number.

With that settled, here is what the sequential depth-first recursion cost.

sequential depth-first    items=68   698.6 ms

Fifteen calls at 45 ms each is 675 ms. Almost exactly the theoretical figure, which means nearly all of that time is spent waiting on round trips. That suggested parallelism would help, so I passed a ThreadPoolExecutor down and called ex.map inside the recursion. That is where it jammed.

recursion inside pool  workers=2  no return after 5 seconds (deadlock)
recursion inside pool  workers=4  no return after 5 seconds (deadlock)
recursion inside pool  workers=8  completed items=68  286.0 ms

Each parent task holds a worker while it waits for its children. With four workers, the root plus three projects fill the pool, and the fourth project queues forever. Nothing raises. It simply never comes back. The run that succeeded at workers=8 only did so because this particular tree happened to fit; add one more project and it stalls the same way.

The cause is recursion sharing a bounded pool, so I dropped the recursion and widened the walk one level at a time instead.

from concurrent.futures import ThreadPoolExecutor
 
 
def fetch_level(store_id: str, prefix: str) -> list[dict]:
    """Fetch exactly one level below prefix, draining pagination."""
    items, cursor = [], None
    while True:
        page = client.beta.memory_stores.memories.list(
            memory_store_id=store_id, path_prefix=prefix,
            depth=1, view="basic", after_id=cursor, betas=BETAS,
        )
        items += [{"path": i.path, "type": i.type} for i in page.data]
        if not getattr(page, "has_more", False):
            return items
        cursor = page.last_id
 
 
def list_level_parallel(store_id: str, prefix: str = "/", workers: int = 8) -> list[str]:
    """Descend the tree level by level. No recursion inside the pool, so it cannot jam."""
    out: list[str] = []
    frontier = [normalize_prefix(prefix)]
    with ThreadPoolExecutor(max_workers=workers) as ex:
        while frontier:
            next_level: list[str] = []
            for page in ex.map(lambda p: fetch_level(store_id, p), frontier):
                for item in page:
                    if item["type"] == "memory_prefix":
                        next_level.append(item["path"])
                    else:
                        out.append(item["path"])
            frontier = next_level
    return sorted(out)

Measured again against the same tree:

Doubling the workers from 8 to 16 does not move the number off 147 ms. This tree is three levels deep — /projects/ to project to subdirectory — and since each level waits on the one above it, nothing gets faster than 3 × 45 ms = 135 ms. The measured 147.8 ms sits just above that floor.

Parallelism buys you the width of the tree, not its depth. The time an audit takes is governed roughly by how many levels it has, not by how many memories live in it. Putting the project name in the first segment and keeping two levels beneath it had, without my intending it, been holding that floor down. It is a useful shape to remember the next time a deeper hierarchy looks tempting.

The 45 ms is a value I injected, not the real latency of the API. What I wanted from this was not the absolute numbers but the shape of the speedup and the position of the ceiling. If you are making the same call for your own setup, measure one round trip first, then map your figure onto this table.

grep will not find all your call sites

By this point you can count what breaks and by how much. What remains is knowing where to fix it. This is the step where I cut a corner and paid for it.

My first move was grep. Finding a path_prefix without a trailing slash felt like a job for one regular expression.

grep -rn 'path_prefix="[^"]*[^/]"' app

Run against my local tree — three files, five call sites — it found exactly one.

app/daily_audit.py:6:        memory_store_id=store_id, path_prefix="/projects/alpha", view="basic")

What it missed were the calls not written as string literals. path_prefix=f"/projects/{name}" is an f-string, so "[^"]*[^/]" never matches it. Neither does a value pulled from a dict. And the longer a codebase runs, the more its call sites drift from fixed strings toward variables. So the bias runs against you: the call sites grep can find are the ones that were easiest to fix anyway.

Walking the syntax tree removes that bias. I used Python's ast to visit every .list(...) call and classify the path_prefix argument as a fixed string, an f-string, or something else.

"""prefix_callsites.py — enumerate memories.list path_prefix arguments with ast,
surfacing calls where the trailing slash cannot be confirmed."""
import ast, pathlib, sys
 
 
def prefix_arg(call):
    for kw in call.keywords:
        if kw.arg == "path_prefix":
            return kw.value
    return None
 
 
def literal_of(node):
    """Return the string for constants, the fixed tail for f-strings, None otherwise."""
    if isinstance(node, ast.Constant) and isinstance(node.value, str):
        return node.value, "literal"
    if isinstance(node, ast.JoinedStr):
        tail = node.values[-1]
        if isinstance(tail, ast.Constant) and isinstance(tail.value, str):
            return tail.value, "fstring"
        return "", "fstring"          # ends in an expression, so no trailing slash
    return None, "dynamic"
 
 
def scan(path):
    tree = ast.parse(path.read_text(), filename=str(path))
    for node in ast.walk(tree):
        if not isinstance(node, ast.Call):
            continue
        if not (isinstance(node.func, ast.Attribute) and node.func.attr == "list"):
            continue
        arg = prefix_arg(node)
        if arg is None:
            continue
        value, kind = literal_of(arg)
        if kind == "dynamic":
            verdict = "REVIEW"        # undecidable statically; a human looks
        elif value.endswith("/"):
            verdict = "ok"
        else:
            verdict = "MISSING /"
        yield path.name, node.lineno, kind, verdict
 
 
if __name__ == "__main__":
    rows = [r for f in sorted(pathlib.Path(sys.argv[1]).rglob("*.py")) for r in scan(f)]
    print(f"{'file':<18}{'line':>5}  {'kind':<8}{'verdict'}")
    for f, ln, kind, verdict in rows:
        print(f"{f:<18}{ln:>5}  {kind:<8}{verdict}")
    bad = [r for r in rows if r[3] != "ok"]
    print(f"\ncall sites: {len(rows)}   needs attention: {len(bad)}")

Here is what it reports on the same three files, running on Python 3.10.12.

file               line  kind    verdict
daily_audit.py        5  literal MISSING /
daily_audit.py        9  fstring MISSING /
daily_audit.py       13  fstring ok
exporter.py           6  dynamic REVIEW
exporter.py          10  literal ok

call sites: 5   needs attention: 3

One from grep, three from the syntax tree. The two it added were the f-string and the dict lookup — both shapes I had actually shipped. Looking back, reaching for grep meant I was only ever searching where searching was easy.

The REVIEW row deserves a note. Line 6 of exporter.py pulls its value from a dict, and both entries in that dict happen to end with a slash, so nothing is broken there. It simply cannot be decided statically. I resisted marking it MISSING because a report that flags things needing no fix is a report nobody opens the following week. Let the tool assert only what it can decide, and say plainly when it cannot decide. I try to hold that line in every check script I write.

One small detail. line is where the method call begins, which sits one line above the number grep reports — the line the argument is written on. The syntax tree says line 5 where grep says line 6. I confused myself once trying to diff the two outputs mechanically, so it seems worth writing down.

Situational guidance

Here are the judgment calls the migration tends to raise, laid out by situation.

One last thing. Right after you bump the beta header, snapshot the total count once. The only reason I caught the halving was that I happened to have the previous day's count on hand. Something that changes without a sound can only be caught by going out and counting it.

Memory changes bite hardest in long-running operations. This path_prefix change cut deepest precisely where the tree had been designed most straightforwardly. If you want to dig into the background of that design thinking, I would also point you to Pitfalls and patterns from running Claude agent memory in production.

I am still feeling my way through much of this area. If this gives even one person who was puzzled by a quiet change in row counts a place to start checking, I would be glad. Thank you for reading.